Does option trading convey stock pric

Jianfeng Hu n

Lee Kong Chian School of Business, Singapore Management University, 50 Stamford

a r t i c l e i n f o

Article history:Received 4 June 2012Received in revised form11 July 2013Accepted 23 July 2013Available online 19 December 2013

JEL classification:

a b s t r a c t

After executing option oaway the underlying stotransactions translates intotal stock order imbalaimbalance independent of options. The analysis shows that the option-induced imbalancesignificantly predicts future stock returns in the cross section controlling for the past stockand options returns, but the imbalance independent of options has only a transitory price

about the underlying stock value.& 2013 Elsevier B.V. All rights reserved.

Subrahmanyam,ks (Chordia and

alternative venue, Easley, O0Hara,

and Srinivas (1998) and Pan and Poteshman (2006) show

option transactions and the underlying stock transactions

Contents lists available at ScienceDirect

journal homepage: www.e

Journal of Financ

Avanidip Baksakimiamitriy

Massachusetts at Amherst, University of Missouri at Columbia, the 2013

Association Conference, the 22nd Annual Derivatives Securities and RiskManagement Conference at the FDIC, and the Sixth Annual Conference on

Journal of Financial Economics 111 (2014) 6256450304-405X/$ - see front matter & 2013 Elsevier B.V. All rights reserved.http://dx.doi.org/10.1016/j.jfineco.2013.12.004and how this interaction affects the predictability of stockreturns.

When customers execute option trades, the optionsmarket makers take the opposite side of the transaction

Asia-Pacific Financial Markets. I owe special thanks to Henry Schwartz forgenerously providing the option tick data at Trade Alert LLC.

n Tel.: 65 6808 5477; fax: 65 6828 0427.E-mail address: jianfenghu@smu.edu.sgthat options order flow also predicts the underlying stockreturns. This paper examines the interaction between

China International Conference in Finance, the Sixth Annual Risk Man-agement Conference in Singapore, the 2012 European Finance Associa-tion Doctoral Tutorial in Copenhagen, the 2012 Northern Financeempirical evidence that supports this predat the market level (Chordia, Roll, and2002) and in the cross section of stocSubrahmanyam, 2004).

The stock options market provides anfor gaining stock exposure. For example

(discussant), Rui Yao, Joe Zhang, Xiaofei Zhao (discussant), Hao Zhou, andthe seminar participants at Baruch College, Central University of Financeand Economics, City University of Hong Kong, DePaul University, DrexelUniversity, Fordham University, Hong Kong University of Science andTechnology, Indiana University, National University of Singapore, RutgersUniversity at Camden, Singapore Management University, Sungkyunk-wan University, Trade Alert LLC, University of Georgia, University of(sellers) cause the stock prices to move up (down). Theiction exists both

Henry Schwartz, Robert Schwartz, Zhaogang Song (discussant), Jun Tu,Mitch Warachka (discussant), Jason Wei, Liuren Wu, Yangru Wu, Shu YanG12G13

Keywords:OptionsOrder flowInformation asymmetryDelta hedgingPrice discovery

I thank G. William Schwert (the editor),(the referee), Linda Allen, Turan Bali, GurdFangjian Fu, Harmeet Goindi, Armen HovJarrow, Mehdi Karoui (discussant), Pete Kyle, D1. Introduction

Microstructure theories such as Ho and Stoll (1981),Glosten and Milgrom (1985), Kyle (1985), and Easley andO0Hara (1987) suggest that a stock0s order flow affects itsprice and that, in particular, transactions initiated by buyers

har Subrahmanyamhi, Jin-Chuan Duan,n, Jian Hua, RobertMuravyev, Lin Peng,G14 impact. Further investigation suggests that options order flow contains important informatione information?$

Road, Singapore 178899, Singapore

rders, options market makers turn to the stock market to hedgeck exposure. As a result, the stock exposure imbalance in optionto an imbalance in stock transactions. This paper decomposes thence into an imbalance induced by option transactions and an

lsevier.com/locate/jfec

ial Economics

longer horizons. This finding suggests that this predict-ability is driven by a permanent information flow insteadof by temporary price pressure. Aggregating the order

J. Hu / Journal of Financial Economics 111 (2014) 625645626and earn the bidask spread. Given the relative scarcity ofoption transactions, the market makers have difficulty inimmediately unloading the positions through trades in theopposite direction. They often have to hold the positionsfor a long time, frequently until the options expire. Toreduce the risk exposure from these positions, a standardpractice of the market makers is to perform delta hedgingby trading on the underlying stocks. Therefore, if optiontransactions generate an imbalance in the stock exposure,the market makers transfer that imbalance to the stockmarket through their delta hedging practice.

To understand the interactions between transactions inthe stock options market and the underlying stock market,this paper decomposes the overall imbalance in stocktransactions into two components: an imbalance inducedby option transactions and the remaining imbalanceinduced by stock market transactions unrelated to optiontrading. To compute the option-induced stock order imbal-ance, the paper assumes that the options market makers(who receive the bidask spread in the transaction) usefull delta hedging and that customers (who initiate theoption transactions by paying the spread) do not hedge.Under these assumptions, the paper computes the option-induced stock order imbalance by aggregating the signedoption transactions weighted by the respective deltaexposure of each option contract.

The direction of each option transaction is determinedaccording to the Lee and Ready (1991) algorithm. Over afixed time interval (say, a day), the volume of the signedoption transactions can be aggregated to generate an orderflow measure for each option contract. Weighting theorder flow of each option contract by its delta exposure,the paper computes a measure for aggregated deltaexposure over all of the option transactions for each stock.This measure provides an estimate of the stock orderimbalance induced by option trading. Subtracting thisoption-induced order imbalance from the total stock orderimbalance results in the residual stock order imbalancethat is unrelated to the option trading activities. This paperanalyzes how these two components interact to predictfuture stock returns.

The paper computes daily stock order imbalance for alarge cross section of stocks with options during the periodof April 2008 to August 2010. On average, the samplecontains 2,207 stocks per day. The paper examines thecross-sectional relation between the two components ofthe stock order imbalance and the stock returns over thenext day, and finds several results new to the literature.First, out of the two components, only the option-inducedorder imbalance positively predicts stock returns over aday controlling for microstructural variables including thepast stock and options returns. Firms in the quintile of thehighest option-induced order imbalance outperform thosein the lowest quintile by 8.74 basis points (bp) on the nextday. This outperformance translates into an annualizedexcess return of 22% with a t-statistic of 6.03. The remain-ing stock order imbalance unrelated to option trading hasno significant forecasting capability for stock returns,although it shows a large contemporaneous price impact.Furthermore, the return predictability from the option-induced order imbalance does not reverse direction inflows over half-hour intervals generates similar patterns inthe predictability of intraday returns.

The finding is particularly interesting because optiontrading generates only a small proportion of the total stockorder imbalance, but this small proportion contains mostof the predictive information about permanent stock pricechanges.1 The finding suggests that the information fromthe options market is not immediately incorporated intothe stock prices. To better understand the meaning andsource of the return predictability, the paper examineshow the return predictability from the option-generatedstock order imbalance varies with option contracts andfirm characteristics.

Subsequently, the paper divides the option-inducedorder imbalance into three groups: for at-the-money(ATM), in-the-money (ITM), and out-of-the-money (OTM)options. Separate return prediction analysis shows that thepredictive power comes mainly from the delta exposuresof the ATM and the ITM options, but not the OTM options.This is possible because most institutional investors usethe OTM options to gain volatility exposure while hedgingthe associated delta exposure by using the underlyingstock. Their hedging activity offsets that of the marketmakers so that the delta imbalance of the OTM optionsdoes not translate into an actual stock order imbalance. Onthe other hand, the investors who have private informa-tion about the stock price movement often take positionsin the ATM or the ITM options.

The paper also divides firms into different groups basedon their characteristics and analyzes the difference inreturn predictability across these groups. First, the firmsare classified into three groups based on their degrees ofinformation asymmetry. The measurement components ofinformation asymmetry are the probability of informedtrading (PIN) proposed by Easley and O0Hara (1987), thenumber of stock analysts following the firm, the bidaskspread in the stock market, the adverse-selection compo-nent of the bidask spread from both Glosten and Harris(1988) and Lin, Sanger, and Booth (1995), and the firm size.Regardless of which measure is used, the option-inducedstock order imbalance always shows the largest predictivecoefficient and t-statistic in the group of informationallyopaque firms, that is, the firms with high PINs, low analystcoverage, large spreads, large adverse-selection compo-nents of the spread, or small market capitalizations. Thesefindings suggest that informed trading in the optionsmarket drives the return predictability from optionsorder flow.

The basis for the second type of classification is theshort-sale constraints. Short-sale constraints are measuredby institutional holdings because firms with more institu-tional holdings are less costly to borrow and sell short(D0Avolio, 2002). The analysis shows that the group offirms with low institutional ownership has both the

1 On average, option trades account for only 3.41% of the total stockorder imbalance.

largest predictive coefficient estimate and the higheststatistical significance for the option-induced order imbal-ance. This finding suggests that option trading, oftenconsidered an effective way to mitigate short-sale con-straints, conveys more information about the firms boundby short-sale constraints. Informed traders with negativenews could place orders in the options market as analternative to short selling, especially when short selling

J. Hu / Journal of Financial Economics 111 (2014) 625645 627is more difficult (Figlewski and Webb, 1993; Johnson andSo, 2012).

The basis for the third type of classification is markettrading activities. The analysis shows that the option-induced order imbalance predicts returns better whenoption trading is active and the total options volume ishigh. This finding is consistent with the prediction ofEasley, O0Hara, and Srinivas (1998) that informed tradersprefer to trade in deep markets.

Further, the paper performs an event study aroundearnings announcements and finds that the option-induced order imbalance predicts the cumulative abnor-mal returns (CARs) five days before the announcement.This predictability is robust to alternative event windowsand becomes greater when the earnings surprise is largeand the dispersion of analysts0 forecasts is high. Overall,the analysis shows that option trading becomes moreinformative when the information asymmetry is severeand the profit from informed trading is high.

The findings in this paper have at least two implica-tions. First, the analysis highlights the importance ofseparating option-induced stock order imbalance fromorder imbalance generated by pure stock market investors.The prior literature often compares order flows and trad-ing volumes from the options and stock markets directly.Through a careful separation, this paper shows thatoptions order flow contains an important informativecomponent about future stock price movement that isnot found in trading activities of pure stock investors.Therefore, this decomposition is critical to measuring theinformation content in order flows.

Second, the findings shed light on the ongoing debateregarding where informed traders place their orders.A number of empirical studies examine the cross-marketinformation flow between the stock and the optionsmarkets.2 In a closely related study, Pan and Poteshman(2006) show that the volume ratio of put and call optionspredicts future stock prices in the cross section. This papergoes a step further and investigates the interaction ofthe two markets via hedging activities. To the best ofmy knowledge, this is the first paper to show that thepermanent price information in stock order flow isinduced mostly by option transactions, which suggeststhat informed traders do use options.

2 See, for example, Manaster and Rendleman (1982), Bhattacharya(1987), Anthony (1988), Stephan and Whaley (1990), Chan, Chung, andJohnson (1993), Easley, O0Hara, and Srinivas (1998), Chan, Chung, andFong (2002), Chakravarty, Gulen, and Mayhew (2004), Cao, Chen, andGriffin (2005), Holowczak, Simaan, and Wu (2006), Pan and Poteshman(2006), Ni, Pan, and Poteshman (2008), and Muravyev, Pearson, andBroussard (2013).The rest of the paper is organized as follows: Section 2reviews the literature on the price impact of optiontrading. Section 3 proposes a decomposition method forthe total stock order imbalance and develops the hypoth-eses for empirical testing. Section 4 describes the dataused in this study. Section 5 reports the empirical resultsof the hypothesis tests. Section 6 provides additionalanalyses on the predictive power resulting from optiontransactions. Section 7 concludes.

2. Background and motivation

Derivative trading can convey important information ina market with information asymmetry. Black (1975) firstnotes the possibility that informed traders could use theoptions market as an alternative trading venue becauseoption contracts provide higher leverage. Biais and Hillion(1994) examine the impact from the introduction ofoptions and find that informed traders0 profits can eitherincrease or decrease depending on the type of liquidityorders. Focusing on how private information gets incorpo-rated into security prices, Easley, O0Hara, and Srinivas(1998) formalize a two-market microstructure model inwhich informed traders choose to trade in the stock oroptions markets. The authors show that a pooling equili-brium exists in which informed traders trade in bothmarkets when the leverage and liquidity of the optionsare sufficiently high. They further argue that because theavailability of multiple option contracts presents difficultlearning problems for uninformed traders, option con-tracts can be more attractive to informed traders. Ifinformed traders trade in the options market, the optionsorder flow can contain information about fundamentalvalues of the underlying stocks.

The options order flow can gain return predictabilityin another important yet less discussed way. Starting withBlack and Scholes (1973) and Merton (1973), the financeliterature has built a modern derivative pricing theory onthe no-arbitrage argument. Other than the price relationbetween options and the underlying security, the no-arbitrage rule also explains how to hedge option positionsby using the underlying security. This type of hedgingresults in a reverse dependence of the stock order flow onthe options order flow. After an options trader executes atransaction, the options market maker takes the oppositeposition and gains risk exposures to the underlying pricemovement as well as other factors such as return volatility.Because the liquidity in the options market is usually nothigh enough for market makers to unload inventoryimmediately (liquidity is further diluted to hundreds ofdifferent option contracts on the same underlying secur-ity), market makers need to hedge the underlying pricerisk by transacting in the underlying market. The hedgeratio is captured by the option price0s sensitivity to theunderlying price movement, delta. Meanwhile, customerscan also perform delta hedging if they intend to gain riskexposures other than the delta. In doing so, they not onlytrade stocks to hedge the delta exposure themselves butalso induce the options market makers to perform deltahedging. Therefore, the options order flow leads to sub-sequent stock order flow through delta hedging transactions.

Overall, no significant return predictability is found fromthe options order flow once the stock order flow is

J. Hu / Journal of Financial Economics 111 (2014) 625645628Even without private information, the hedging transactionscan cause transitory pressure on the stock price in the sameway as stock transactions unrelated to the options market.Without perfect liquidity, such price pressure drives thestock price away from its fundamental value temporarily,followed by a reverse price movement to the fundamentalvalue later. Empirically, Ni, Pearson, and Poteshman (2005)show that the practice of delta hedging causes the under-lying stock prices to cluster around the option strike priceson option expiration dates. In the structured equity productmarket, Henderson and Pearson (2010) find that hedgingtransactions raise the underlying stock price by almost onehundred bp on the pricing date.

The theoretical discussion on the feedback effect fromoptions has motivated many empirical studies on the lead-lag relation between the two markets. A group of research-ers focus on the relation between the actual stock pricesand the implied stock prices from options. Early studiessuch as Manaster and Rendleman (1982) and Bhattacharya(1987) find that the options market leads the stock marketin price discovery. However, more evidence points in theopposite direction. For example, Chakravarty, Gulen, andMayhew (2004) find that the information share of optionquotes is less than 20% on average. Holowczak, Simaan,and Wu (2006) show that the information share of optionseven decreases over time, and they argue that this decreaseis because of the prevailing use of computers for the auto-matic updating of quotes in the options market. In a recentstudy, Muravyev, Pearson, and Broussard (2013) find thatoption quotes, but not stock quotes, adjust themselves toeliminate arbitrage opportunities across the two markets,and they conclude that the options market does not play arole in price discovery. Another strand of empirical researchdirectly investigates the options order flow. Many studiesfind that this order flow predicts future stock returnsin time series regressions for a small group of stocks, forexample, Easley, O0Hara, and Srinivas (1998), Poteshman(2006), Dong and Sinha (2011), and Holowczak, Hu, and Wu(2014).

Pan and Poteshman (2006) investigate the informationcontent of the options order flow in the cross section ofstocks. Using a unique data set, these authors constructvolume ratios of put and call options from buyers openingnew positions. They find that the open-buy put-call rationegatively predicts returns and that the return predictabil-ity lasts over three weeks. However, most studies on theoptions order flow do not control for the stock order flow.It is still unclear how options and stock order flows interactto generate return predictability. Because informed traderscan trade in both markets, it is important to understandwhether the options market makes a marginal contributionto embedding private information in stock prices.

Chan, Chung, and Fong (2002) investigate the informa-tional role of order flows and quote revisions in the twomarkets. They find that the options order flow does nothave a pricing effect but the stock order flow does. Cao,Chen, and Griffin (2005) examine order flow informationaround mergers and acquisitions. Their results suggestthat the options volume imbalance becomes significantlyinformative about future stock prices right before mergerannouncements but remains silent during normal periods.controlled for. However, the order imbalances analyzedin these two studies come directly from trading volumes inthose two markets without separating out the hedgingtransactions. Thus, the possibility exists that the contami-nated total stock order imbalance absorbs both the infor-mation content and the return predictability from theoptions order flow. Therefore, the information content indifferent markets must be separated carefully before anylead-lag analysis can be conducted.

This paper is also motivated by some recent develop-ments in the options market. Most studies use data prior to2000 when the options market in the United States was notyet a consolidated national market (see Battalio, Hatch, andJennings, 2004). Multiple listings on different exchangesand price competition were not common at that time,and transaction costs in the options market were very high.Since 2000, as a result of Securities and ExchangeCommission0s efforts and the development of quote updat-ing and order routing technologies, the options market hasexperienced tremendous growth and consolidation. Marketcompetition drives down transaction costs and improvesliquidity. For informed traders, these changes are likely toaffect their decisions on order allocations. Therefore, it isworthwhile to reexamine the information content in theoptions order flow in a more recent sample period. Also, thelack of liquidity in the past significantly limits the samplesize for an empirical analysis. For example, Chan, Chung,and Fong (2002) have only 14 stocks in their sample inwhich the stock with the most active option trading,General Electric (GE), has an average daily trading volumeof 2,595 option lots. By comparison, GE0s options volumereached 134,874 lots on April 1, 2008, the first day in mysample and the average options volume of all of thecommon stocks is 3,255 lots on that day. The growth ofthe options market significantly facilitates the expansion ofthe sample coverage. This study covers all of the commonstocks with options and focuses on the cross-sectionalrelation between the order flows and the future returns.

3. Methodology

This section explains the calculation of the orderimbalances in the options and the stock order flows. Itthen describes the primary hypotheses and the empiricaltesting methods.

3.1. Options order imbalance

A remarkable feature of the options market is thathundreds of option contracts underlie the same security.For example, on January 2, 2009, there were 478 optioncontracts underlying Apple0s common stock that spanned61 strike prices from $12.5 to $400 and six expiration datesfrom 15 days to 750 days.3 Because each contract has itsown unique combination of put or call type, strike price,and maturity, the trading process for each option contract

3 The closing price of Apple0s stock was $90.75 on January 2, 2009.

discloses the underlying price information in differentways. Measuring the price information in the optionsorder flow is challenging. In a recent study, Holowczak,Hu, and Wu (2014) show that risk exposure to the under-

equation for the cross section of stocks:

Reti;t 5

k 1k1 SOIi;tk

5

k 1k2 OOIi;tkXi;t1i;t ;

3where, for stock i on day t, the Reti;t is the stock returncalculated using the midpoint of the last National Best Bidand Offer (NBBO) prices before the market close.5 Further-

J. Hu / Journal of Financial Economics 111 (2014) 625645 6293.2. Stock order imbalance

Because options market makers take the opposite sideof the OOI, the OOI also measures the net delta hedgingdemand in the stock market. Then the stock order imbal-ance (SOI) unrelated to the options market is the differ-ence between the total order imbalance (TOI) in the stockmarket and the OOI. The SOI is defined as

SOIi;t TOIi;tOOIi;t Nj 1Diri;t;j sizei;t;j

Num_shares_outstandingiOOIi;t ;

2where TOIi;t is in terms of the trading volume, and Diri;t;jand sizei;t;j are the direction dummy and the size of the jthstock transaction of firm i on day t, respectively. The TOI isalso scaled by the number of shares outstanding to beconsistent with the OOI. After subtracting the OOI from theTOI, the remainder is the order imbalance caused by stockmarket investors.

3.3. The main test

To gauge the return predictability from the order flowsin the two markets, the paper estimates the following

4 Experiments with alternative scalers such as total delta volume andtotal stock volume produce similar results. But these alternative scalersare more likely to generate outliers during inactive trading periods.Hence, I do not report those results in the paper.OOIi;t nj 1100Diri;t;j deltai;t;j sizei;t;j

Num_shares_outstandingi: 1

For stock i on day t, the Diri;t;j is a dummy variable equal toone (negative one) if the jth option trade is initiated bythe buyer (seller) according to certain trade signing algo-rithms. The deltai;t;j is the option price0s sensitivity to theunderlying stock price, and the sizei;t;j denotes the tradesize in option lots (one hundred shares of the underlyingstock). The numerator is thus the aggregate delta positionof the active options traders who initiate the trades.The net delta exposure captures the directional tradingintention about the underlying stock price, which disclosesthe price information in the options market if any. Theorder flow is more likely to be imbalanced when tradingis inactive, and this imbalance can merely reflect noisefrom uninformed trades. To address this issue and also forcross-sectional standardization, the final OOI is the netdelta exposure scaled by the number of common sharesoutstanding.4lying stock price (delta) is an important considerationwhen extracting the stock price information from optiontransactions. Following the authors0 reasoning, I propose astandardized measure of the options order imbalance(OOI): more, Xi;t1 is a set of control variables composed of the

closing percentage bidask spread of the stock, the stock0sturnover ratio as the total trading volume scaled by thenumber of shares outstanding, the log trading volumes inthe two markets, and the stock returns and the equallyweighted options returns from the past five days.6

Hypothesis 1. The SOI positively predicts the future stockreturns, at least for some k, k140. If the predictabilitycomes from the price pressure of liquidity trades, then thepredictive relation reverses its sign in longer horizons. Ifthe predictability comes from informed trading in thestock market, then the predictive relation does not reverseits sign.

Informed trading in the stock market allows the SOIto predict the permanent price changes (see, e.g., Glostenand Milgrom, 1985; Kyle, 1985; Easley and O0Hara, 1987).A positive (negative) SOI reflects the buying (selling) ofinformed traders with good (bad) news and should predicta higher (lower) stock price in the future. Alternatively,even in the absence of information asymmetry, the stockmarket0s liquidity providers can lift the price quote afterreceiving a buy order and reduce the price quote afterreceiving a sell order to entice the offsetting orders andreturn to their optimal portfolios. As a result, the SOI alsopositively predicts future price changes. However, thisprice impact is short-lived. Once the price pressure fromthe order imbalance disappears, the stock price returns toits fundamental value. Thus, a reverse (negative) predictiverelation exists in longer horizons.

Hypothesis 2. The OOI positively predicts the future stockreturns controlling for the SOI, at least for some k, k240.If the predictability comes from the price pressure of theoptions market makers0 delta hedging, then the predictiverelation reverses its sign in longer horizons. If the predict-ability comes from informed trading in the options market,then the predictive relation does not reverse its sign.

If informed traders gain stock exposures in the optionsmarket, then the OOI captures that information content.Similar to the SOI, a positive (negative) OOI reflects theprivate information about good (bad) news and predicts thatthe underlying stock price will go up (down) permanently

5 In theory, the private information in the order flows is aboutspecific firms, and the imbalance should have a greater predictive abilityabout the idiosyncratic returns. In unreported tests, I use the risk-adjusted returns as the dependent variable and find stronger results.I have also experimented with open-to-close NBBO midpoint returns andthe returns calculated by using transaction prices. The results are largelythe same.

6 I have also experimented with the value-weighted options returnsand found similar results.

and high) with cutoff points at the 30th and 70th percen-

trade options only. Institutional ownership as a proxy for

options volume stocks.

J. Hu / Journal of Financial Economics 111 (2014) 625645630after that information becomes public. Meanwhile, theoptions market makers perform delta hedging to reducetheir risk exposures and transfer the OOI to an imbalancein the stock market. The OOI also causes price pressure onthe underlying stock, but the effect is temporary. Becausethe private information and the price pressure can comefrom both markets, understanding the OOI0s marginalcontribution to the price discovery is important. Eq. (3)investigates the return predictability from the OOI con-trolling for the SOI.

3.4. Conditional tests

If the OOI does predict future stock returns, it would beinteresting to investigate how the predictability dependson option contract types and firm types. The followingtests focus on the sources of the potential return predict-ability from the OOI and serve as robustness checks on theinformational role of the OOI.

3.4.1. Option leverage

Hypothesis 3. If informed traders use options for leverage,they should prefer options that give them the highestleverage. Therefore, the order imbalance of these optionsis more informative about future stock returns.

To test this hypothesis, the paper divides all of the optiontransactions into three groups based on the moneyness of theoption, delta: out-of-the-money (OTM, jdeltajo0:375), at-the-money (ATM, 0:375r jdeltajr0:625), and in-the-money(ITM, jdeltaj40:625). Although the OTM options have thesmallest delta, they provide the highest leverage for everydollar invested because of their low prices. The ITM optionshave the lowest dollar leverage among the three groups. Icalculate the OOI separately for the three option groups andestimate the following equation:

Reti;t 5

k 1k1 OTM_OOIi;tkk2 ATM_OOIi;tk

k3 ITM_OOIi;tk 5

k 1k4 SOIi;tkXi;t1i;t ; 4

where Reti;t , SOIi;tk, and Xi;t1 are the same as definedbefore; and OTM_OOIi;tk, ATM_OOIi;tk, and ITM_OOIi;tkare delta order imbalances calculated using the OTM, theATM, and the ITM options, respectively.

3.4.2. Level of information asymmetry

Hypothesis 4. If the options order flow discloses privateinformation, then the predictive ability of the OOI shouldbe greater for the firms with more information asymmetry.

The likelihood of information-based trading is higherwhen more private information exists to be exploitedceteris paribus. For less transparent firms, the informationcontent in the OOI can be higher and, thus, the returnpredictability can also be stronger. The paper uses fiveproxies for the level of information asymmetry: the PINfrom Easley and O0Hara (1992), the number of analystsfollowing the firm, the percentage bidask spread, theadverse-selection component of the bidask spread, andThe same return prediction analysis is repeated for thethree groups based on the daily option trading volume,which is a proxy for the options market liquidity.

4. Data

Analysis of the order flows in both markets requires themerging of several databases. This section describes thedetails of the data sources, the sample selection, and thevariable construction.the market supply of short interests has a negative relationto the difficulty of selling short (D0Avolio, 2002; Asquith,Pathak, and Ritter, 2005). But, institutions are also morelikely to have an information advantage over retail inves-tors. Thus, the paper proposes Hypothesis 5.

Hypothesis 5. A stock0s institutional ownership can causethe OOI0s predictive ability to decrease because the insti-tutional ownership reduces the short-sale costs. Alterna-tively, a stock0s institutional ownership can cause the OOI0spredictive ability to increase because the institutionalinvestors could be better informed.

Similar to the tests discussed in Section 3.4.2, Eq. (3) isreestimated for the three groups based on the institutionalownership: low (o30%), medium (3070%), and high(Z70%). The analysis then compares both the statisticaland economic significance of the estimated OOI coeffi-cients for the groups.

3.4.4. Market activityThe liquidity in the options market can also affect the

predictive ability of the OOI if the return predictability iscaused by informed trading.

Hypothesis 6. The price impact of the OOI increases whenthe liquidity in the options market improves because theinformed traders can conceal their trades better in hightiles. I then reestimate Eq. (3) for each group and comparethe estimated coefficients and t-statistics of the OOI acrossgroups.

3.4.3. Institutional ownershipOther than the level of information asymmetry, the

short-sale costs also have an impact on the predictiveability of the OOI. In practice, options frequently are usedas devices to circumvent the short-sale constraint in thestock market. For stocks more difficult to sell short, optiontrading can have greater informational benefits becauseinformed traders with negative news can be forced tothe firm size. To calculate the adverse selection componentof the spread, the paper uses both the Glosten and Harris(1988) and Lin, Sanger, and Booth (1995) models.

Based on each proxy of information asymmetry, thepaper first sorts all firms by ascending order every day anddivides the full sample into three groups (low, medium,

Augushour tPricede prs: buyof firmng stok-Schptions

Buy c

1,31,71,01

1774,1

,314,5254

J. Hu / Journal of Financial Economics 111 (2014) 625645 631Table 1Summary statistics of the options market.The table describes the options market activity between April 2008 and

days are included in the sample. The following trades are excluded: (i) off-close (the last five minutes), (iii) trades that are reported to the Optionsstructured trade, and (v) data errors such as zero strike prices or zero tratransaction-type groups based on the trade directions and the option typeand Ready (1991) algorithm. Panel A provides the statistics for the numbercounted as option lots (one lot equals one hundred shares of the underlyitotal volume by the option moneyness, delta, and the maturity, T. The Blacand a zero dividend rate. Panel D gives the volume breakdown by the omoney.

Statistics All

Panel A: Number of firms per dayMean 1,670Maximum 1,909Minimum 1,395Standard deviation 98

Panel B: Transaction statistics

Mean trade size 17.45Mean daily number of trades 273,102Mean daily trade volume 4,765,903 1Mean daily premium (in millions) 1,031.364.1. Options market activity

The options transaction data of 611 trading days in theperiod of April 2008 to August 2010 come from Trade AlertLLC, a specialized data vendor for the options market.The data comprise all of the trade messages recorded bythe Options Price Reporting Authority (OPRA), a nationalinformation processor that consolidates market informa-tion generated by option trading on all of the US optionsexchanges. In real time, Trade Alert matches each optiontransaction record with the underlying stock price andcomputes the implied volatility from a binomial tree.Following the quote rule, Trade Alert also classifies anoption trade as buyer-initiated (seller-initiated) if thetransaction price is above (below) the most recent midquote price. Such comprehensive trade data enable across-sectional study over a relatively long period.

The quote rule fails to classify a trade when the tradeprice falls at the mid quote or when no valid quote exists.In such cases, I apply the tick rule: If the trade price isabove (below) the last different trade price, it is classifiedas buyer-initiated (seller-initiated). This procedure of applying

Panel C: Percentages of trading volume by moneyness and maturity

OTM: jdeltajo37:5% 51.94 13ATM: 37:5%r jdeltajr62:5% 33.91 10ITM: jdeltaj462:5% 14.14 410oTr30 days 36.64 1031oTr60 days 27.91 7T460 days 35.45 9

Panel D: Percentages of trading volume by exchange

American Stock Exchange 7.15 1Better Alternative Trading System 0.05 0Boston Options Exchange 4.70 1Chicago Board Options Exchange 29.66 8International Securities Exchange 27.43 7Nasdaq 2.50 0New York Stock Exchange ARCA 11.59 2Philadelphia Stock Exchange 16.92 4t 2010. Only single-name equity options with maturities of more than tenrades, (ii) trades at the market open (the first 15 minutes) and the marketReporting Authority as late or cancel, (iv) trades flagged as part of aices. All statistics are reported for the full sample as well as for the fourcall, sell call, buy put, and sell put. The trade direction is based on the Lees per day. Panel B reports the trade level statistics. The trading volume is

cks). Panel C shows the trading volume breakdown as percentages of theoles-Merton model is used to calculate the delta with a zero interest rateexchange. OTMout-of-the-money; ATMat-the-money; ITM in-the-

all Sell call Buy put Sell put

90 1,400 1,179 1,16409 1,738 1,634 1,50391 1,086 839 87420 115 120 102

.74 16.16 18.80 17.6714 75,321 49,145 47,14726 1,217,082 924,072 832,880.07 246.89 223.15 206.45the tick rule after the quote rule first appears in Lee andReady (1991).

The analysis excludes all indexes, units, AmericanDepositary Receipts (ADRs), Real Estate Investment Trusts(REITs), closed-end funds, exchange-traded funds (ETFs),and foreign firms and focuses on common stocks only.For computing the OOI, I exclude the options that expirewithin ten calendar days and the following trades:(i) off-hour trades, (ii) trades at the market open (the first15 minutes) and the market close (the last five minutes),(iii) trades that are reported to the OPRA as late orcancel, (iv) trades flagged as part of a structured trade,and (v) data errors such as trades with zero prices or zerostrike prices. I filter out (i)(iii) because the trade directionclassification is less reliable for those trades; (iv) becausethese trades are less likely to contain stock price informa-tion than volatility information; and (v) for obviousreasons. Also excluded are the days with a stock-split orwith dividends because of their complex implications foroption pricing and trading.

Table 1 presents the primary statistics of the fullsample of options as well as those of four transaction-type

.21 11.77 11.62 10.21.11 9.52 5.75 5.21.26 4.25 2.03 2.06.25 9.22 7.12 6.46.78 7.05 5.46 4.82.55 9.27 6.81 6.20

.93 1.82 1.32 1.28

.01 0.01 0.01 0.01

.24 1.17 1.01 0.97

.76 8.05 5.79 5.06

.06 6.88 5.58 4.97

.64 0.65 0.45 0.44

.69 2.62 2.08 1.90

.79 4.40 3.22 2.91

groups: buy call, sell call, buy put, and sell put. Sometrades do not fit into any of the four transaction-typegroups. I do not report statistics for these trades becausethey are not of interest to the study.7 On average, 1,670stocks per day have valid option trades, with the max-

J. Hu / Journal of Financial Economics 111 (2014) 625645632imum at 1,909 and the minimum at 1,395. The call optionsare traded more often than the put options because theaverage daily number of firms with call transactionsexceeds the average daily number of firms with puttransactions in both the buy and the sell categories inPanel A. Trade level statistics are reported in Panel B. Theaverage trade size is 17.45 lots in the full sample. The calloption trades are, on average, smaller than the put optiontrades. However, the call options have a much largeraverage daily number of trades than the put options.Therefore, the average daily volume of the call optionsalso exceeds that of the put options by approximately 0.77million lots. The average premium of all single-nameoptions on a day exceeds $1 billion. Lakonishok, Lee,Pearson, and Poteshman (2007) find that non-market-maker participants take net long positions in optiontrading volumes from 1990 to 2001. Panel B confirms theirfinding by showing that for both call and put options, thereare more buy trades than sell trades in terms of bothvolume and premium. Panel C presents the distributionof the trading volumes (lots of options traded) across theoption moneyness and the time to expiration. Across themoneyness, the OTM options are the most frequentlytraded, accounting for over 50% of the entire market0svolume; the ATM options account for 33.91% of the totalvolume; and the ITM options have the smallest share, withonly 14.14%. The same pattern exists for the four transaction-type groups. In addition, across the three moneyness regions,the non-market-maker participants take net long positionsexcept for the ITM puts. The difference between the buy andsell volumes is the largest for the OTM options and thesmallest for the ITM options. Across maturities, approxi-mately 37% of the option transactions expire in 30 calendardays. The options that expire in 3160 days account for27.91% of the total volume. Given the fact that the analysisexcludes all options expiring in less than ten days, mosttrades clearly are short-term options. Again, the samepattern exists for the four transaction-type groups, and thebuy volume always exceeds the sell volume.

At the beginning of the sample period, there wereseven options exchanges in the OPRA plan: the AmericanStock Exchange (AMEX), the Boston Options Exchange(BOX), the Chicago Board Options Exchange (CBOE), theInternational Securities Exchange (ISE), Nasdaq, the NewYork Stock Exchange ARCA (NYSE), and the PhiladelphiaStock Exchange (PHLX). During the sample period, anotherparticipant exchange, the Better Alternative Trading Sys-tem (BATS), joined the OPRA on February 26, 2010. The lastpanel of Table 1 presents the market share by exchanges.The CBOE and the ISE lead in market share, as each of

7 The trade direction is assigned to zero for these trades. Therefore,unclassified trades have no impact on the OOI in the empirical analysis.However, omitting these trades can make a full-sample statistic differentfrom the sum or weighted average of the numbers in the four transaction-type groups.them attracts more than one quarter of the total volume:29.66% and 27.43%, respectively. The PHLX is also compe-titive, attracting 16.92% of the total volume, followed bythe NYSE at 11.59% and the AMEX at 7.15%. The BOX andNasdaq are relatively small markets; neither has morethan 5% of market share. The BATS does not have much ofa market share because it is newly established. TheHerfindahl index of the market share is 0.213, indicatingthat the US options market is moderately concentratedduring the sample period.

4.2. Other data sources

The stock transaction data come from the NYSE Tradeand Quote (TAQ) database. I follow Lee and Ready (1991) toassign stock transaction directions. However, unlike Leeand Ready (1991), I collapse the trades that happen in thesame second into one record weighted by the dollarvolume and use the NBBO prices one second before thetrade time (t1) for the trade signing. In addition toexcluding canceled trades and data errors, the analysisexcludes trades within the first 15 minutes and the lastfive minutes of trading hours to increase the accuracy oftrade signing and to match the observation period ofthe OOI.

The stock returns are calculated using the midpoints ofthe closing bid and ask prices collected from the Center forResearch in Security Prices (CRSP). The number of analystsfollowing is extracted from the Institutional Brokers0

Estimate System (I/B/E/S), and the institutional ownershipdata are from the Form 13-F filings in the Thomson Reutersdatabase.

To be included in the final sample, a stock-day observa-tion must have valid stock price information from the CRSPwith share code 10 or 11 and the stock must have optionslisted on an options exchange in the US. On average, thefinal sample has 2,207 stocks a day.

4.3. Statistics of the main variables

The OOI and the SOI are constructed as discussed inSections 3.1 and 3.2. The option delta is calculated basedon the Black and Scholes (1973) model assuming a zerointerest rate and a zero dividend rate for simplicity. If nooption (stock) transaction exists on a particular day, theOOI (SOI) is set to zero. Panel A of Table 2 presents the timeseries averages of the cross-sectional statistics for the mainvariables. For ease of reporting, I scale the stock returnsand all order imbalance variables by basis points. Theaverage TOI is close to zero (0.782 bp) with the medianeven smaller at 0.348 bp. The standard deviation of theTOI reaches 21 bp, which is approximately 27 times itsmean. As a result, the TOI is not significantly different fromzero, which indicates that the stock market is balancedoverall. In extreme cases, the maximum imbalance goesbeyond 2.7% of all shares outstanding and the minimum isbelow 3.98%. The SOI has similar statistics to those of theTOI with a mean of 0.773 bp and a standard deviation of21.71 bp. The OOI is even smaller, on average, with a meanof 0.016 bp but the standard deviation is not that small(5.517 bp). The skewness of the OOI (0.272) is also much

et is tstocknce. Sf shar5 p.mcorrel

ian

05484301

31450383

I

J. Hu / Journal of Financial Economics 111 (2014) 625645 633Table 2Descriptive statistics of the main variables.Panel A reports the time series averages of the cross-sectional statistics. R

Best Bid and Offer (NBBO) prices before the market close. TOI is the totaltrading. OOI is the options order imbalance measured as the delta imbalamarket. Turnover is the total stock trading volume scaled by the number ototal stock volume and the log total options volume from 9:45 a.m. to 3:5presents the time series averages of the contemporaneous cross-sectionalfive lags for each variable.

Panel A: Descriptive statistics

Variable N Mean Std Med

Ret 1,348,555 4.602 407.019 6.0TOI 1,348,555 0.782 21.178 0.3SOI 1,340,407 0.773 21.710 0.3OOI 1,340,407 0.016 5.517 0.0Spread 1,348,555 0.279 0.636 0.1Turnover 1,348,552 1.375 1.923 0.9Voloptions 1,348,555 7.900 4.593 9.0Volstock 1,346,313 13.195 1.558 13.0

Panel B: Correlations

Ret TOI SOI OOsmaller than that of the SOI (3.058). Compared with thestock market, the options market is a more balancedmarket. Although the tails of the OOI are not as long asthose of the SOI, the OOI has a larger excess kurtosis (327)than the SOI (182), indicating that both have fat tails. Theaverage size of the bidask spread is 0.28% of the midpointprice in the stock market. It is interesting to comparethe imbalance variables with the turnover ratio becausethe turnover ratio is also scaled by the number of sharesoutstanding. The average turnover ratio is approximately1.38%. At the ratio0s maximum, over 46% of the sharesoutstanding change hands on a given day. The stock andoptions volumes are reported as log total volumes traded.The average log stock volume is around 13.2, and theaverage log options volume is 7.9. It is obvious that theoptions market is still not the same size as the stockmarket.

Panel B reports the time series averages of the dailycross-sectional correlations. The stock returns have largeand positive contemporaneous correlations with all of theorder imbalance variables, and the correlation with the TOIreaches 0.211. These positive correlations suggest that the

Ret 1TOI 0.211 1SOI 0.188 0.962 1OOI 0.082 0.067 0.188 1Spread 0.007 0.040 0.040 0.000Turnover 0.068 0.072 0.071 0.001Voloptions 0.027 0.019 0.019 0.000Volstock 0.025 0.008 0.006 0.005

Panel C: Autocorrelations

Lag Ret TOI SOI OOI

1 0.021 0.105 0.101 0.0282 0.031 0.053 0.051 0.0093 0.016 0.035 0.034 0.0064 0.018 0.028 0.028 0.0045 0.028 0.026 0.026 0.003he daily stock return calculated by using the midpoint of the last Nationalorder imbalance. SOI is the stock order imbalance independent of optionpread is the bidask spread as percentage of the mid quote in the stockes outstanding, reported as percentages. Volstock and Voloptions are the log., respectively. Ret, TOI, SOI, and OOI are reported as basis points. Panel Bations. Panel C gives the cross-sectional averages of autocorrelations up to

Minimum Maximum Skew Kurt

3269.897 7193.655 2.698 111.010397.915 269.771 3.276 181.018403.796 282.378 3.058 182.394106.223 109.196 0.272 327.1250.550 14.153 9.929 180.0130.018 46.133 10.558 238.3200.000 17.667 0.658 0.6637.623 19.464 0.249 0.149

Spread Turnover Voloptions Volstockimbalances have a strong impact on the contemporaneousstock prices. The SOI has a larger contemporaneous corre-lation with the returns (0.188) than the OOI does (0.082).The correlation between the TOI and the SOI is 0.962, butthe correlation between the TOI and the OOI is only 0.067,suggesting that the majority of the TOI is determinedby the stock market investors. The correlation betweenthe SOI and the OOI is negative (0.188); 50.63% of theobservations show the SOI and the OOI with differentsigns. The negative correlation could be because of anunderestimation of structured trades in the sample, suchas the covered calls and the protected puts. For thosetrades, the customers0 hedging transactions in the stockmarket take the opposite direction of their stock exposurein the options market. If the independent stock transac-tions are perfectly balanced, then the SOI has a differentsign from the OOI as a result of the structured trading. TheOPRA flags the structured trades when the customersdisclose such information to their brokers, and all of thesetrades are excluded in the analysis. However, if customersdesign the structured trades on their own, then thesetrades cannot be identified. Thus, they weaken the return

10.034 10.258 0.325 10.233 0.434 0.674 1

Spread Turnover Voloptions Volstock

0.170 0.436 0.322 0.3580.157 0.309 0.265 0.2950.152 0.253 0.233 0.2560.153 0.226 0.215 0.2390.152 0.207 0.202 0.219

75th percentile 0.028 0.174

J. Hu / Journal of Financial Economics 111 (2014) 625645634predictability by adding noise to the imbalance measures.The imbalance variables have small correlations with thespread, the turnover, and the trading volumes. The stockmarket and the options market tend to be active at thesame time because the correlation between the stock andthe options volumes is 0.674.

Panel C presents the cross-sectional averages of theautocorrelations for each variable up to five lags. Theautocorrelations of the TOI and the SOI are strong andpositive. For example, the autocorrelation of the TOI is0.105 at the first lag, and it gradually decays to 0.026 at thefifth lag. By comparison, the autocorrelation of the OOI ismuch smaller. It is only 0.028 at the first lag and almostdies out at the fifth lag. The bidask spread, the turnoverratio, and the stock and options volumes all have highlypositive and persistent autocorrelations.

The relation between the OOI and the SOI is furtherexamined in Table 3. This table shows the statistics ofthe two OOI-to-SOI ratios: one original and one using theabsolute values. To address the ratio explosion when theSOI is close to zero, I winsorize both ratios at the 0.5% and

90th percentile 0.202 0.62999th percentile 3.088 6.388Maximum 25.094 35.898Mean 0.028 0.341Standard deviation 0.986 1.360Table 3Comparing the options order imbalance and the stock order imbalance.This table compares the magnitude of the options order imbalance

(OOI) and the stock order imbalance (SOI). Two ratiosone original andone using absolute valuesare calculated in the full sample, and detailedstatistics are presented. Both of the ratios are winsorized at 0.5%and 99.5%.

Statistic OOI/SOI jOOI=SOIj

Minimum 19.689 01st percentile 3.354 010th percentile 0.309 025th percentile 0.040 0.002Median 0.000 0.03399.5% levels. The original OOI-to-SOI ratio has a meanof 0.028. The standard deviation of the ratio is 0.986.The average absolute ratio is 0.341, but the median is only0.033. Even the 90th percentile absolute ratio is smallerthan one (0.629). Clearly, the OOI is small compared withthe SOI.

5. Empirical results

This section tests the hypotheses in Section 3.

5.1. Predicting stock returns using the SOI and the OOI

The empirical analysis investigates the predictive abil-ity of the SOI and the OOI for stock returns as detailed inSection 3.3. Table 4 contains the main results of the Famaand MacBeth (1973) regressions. The table reports theaverage slope coefficients estimated from the cross-sectional regressions and the Newey and West (1987)adjusted t-statistics. The demeaned coefficient estimatesshow little autocorrelation. For many series, the autocor-relation estimates are insignificant even at the first lag. Themaximum length of a significant lag for all series is three.To be conservative in the reported t-statistics, I chooseto use three lags for all of the Newey and West standarderror calculations. In preparation for the full model test ofEq. (3), I test the predictive ability of the TOI in the firstcolumn by estimating the following equation:

Reti;t TOIi;t1i;t : 5The TOI has an average slope coefficient of 0.002, andthe t-statistic is 0.05. These values suggest that the TOIhas no significant predictive power. Column 2 presents theregression results from using decomposed order imbal-ances. The OOI coefficient of 0.59 is statistically significantat the 1% level (t-statistic5.59). The SOI coefficient ispositive but insignificant (t-statistic0.54). Only the OOIpositively and significantly predicts the next day0s returns.I then add the lagged order imbalance variables to themodel to investigate whether the predictive relationsreverse in longer horizons. Column 3 shows that the SOIon day t2 has a significant and negative coefficient of0.106 (t-statistic2.66), but the lagged OOIs have onlyinsignificant coefficients. The full model in Eq. (3) isexamined in Column 4. With the microstructure controlvariables, the OOI still significantly predicts the next day0sreturns (t-statistic4.17). The reversal effect of the SOIbecomes less significant as the t2 SOI coefficient dropsto 0.065 (t-statistic1.9). I also find that both of thestock and options trading volumes significantly predictreturns, but in opposite directions. A large stock volume isassociated with positive future returns, but a large optionsvolume is associated with negative returns. This findingis consistent with Johnson and So (2012), which showsoptions-to-stock volume ratio negatively predicts stockreturns.

Some concerns might exist that the sample periodcovers the major financial crisis in 2008 when unusualmarket volatility and frequent regulatory interventioncould have distorted the generality of the results. For arobustness check, the full model is reestimated using datafrom 2009 and 2010 only. The results are reported inColumn 5. Comparing Columns 4 and 5, I find that theresults are similar, except that, in the after-crisis period,the first lag of the SOI becomes marginally significant.To make sure that the results are not driven by the choiceof the time window for constructing the order imbalances,I also measure the SOI and the OOI at one hour before themarket close and repeat the full regression in Column 6.The results are largely the same.

The results in this table have important implications forthe first two hypotheses. On the one hand, the SOI doesnot positively predict future returns, although Table 2shows that the contemporaneous correlation is strong.The reversal effect on day t2 indicates that the priceimpact of the SOI is more likely to come from transitoryprice pressure rather than from an information flow. Onthe other hand, the OOI positively predicts the stockreturns on the next day, and the predictability does notreverse in longer horizons. These findings thereby provide

Table 4Daily regressions of stock returns on lagged stock and options order imbalances.The first column reports the Fama and MacBeth regression results of the following equation:

Reti;t TOIi;t1i;t :Reti,t is stock i0s return on day t calculated by using the midpoint of the last National Best Bid and Offer (NBBO) quote before the market close. TOIi,t1 isstock i0s total order imbalance on day t1. The rest of the columns present the Fama and MacBeth regression results based on the following equation:

Reti;t 5

k 1k1 SOIi;tk

5

k 1k2 OOIi;tkXi;t1i;t :

SOIi,tk is stock i0s order imbalance independent of option trading on day tk. OOIi,tk is the options order imbalance measured as the delta imbalance.Xi,t1 is a set of control variables on day t1, including Reti,tk, the stock returns for the previous five days; OpReti,tk, the equally weighted optionsreturns across all of the option contracts on stock i for the previous five days; Spread, the percentage stock bidask spread; Turnover, the ratio of total stocktrading volume to the number of shares outstanding; Volstock, the log stock volume from 9:45 a.m. to 3:55 p.m.; and Voloptions, the log options volume from9:45 a.m. to 3:55 p.m. Column 5 gives the regression results in the subsample period from 2009 to 2010. Column 6 reports the regression results by usingorder imbalances measured at 3 p.m. every day. Standard errors are calculated with the Newey and West adjustment to three lags. The resulting t-statisticsare reported in parentheses. nnn, nn, and n denote statistical significance at the 1%, 5%, and 10% level, respectively.

Variable (1) (2) (3) (4) (5) (6)

Intercept 4.957 3.978 3.899 35.116nnn 26.165nn 29.801nn(0.51) (0.41) (0.40) (2.90) (2.09) (2.53)

TOIt1 0.002(0.05)

SOIt1 0.023 0.037 0.019 0.073n 0.120nnn

(0.54) (0.85) (0.48) (1.65) (2.73)SOIt2 0.106nnn 0.065n 0.054n 0.064n

(2.66) (1.90) (1.67) (1.72)SOIt3 0.046 0.034 0.027 0.045

(1.02) (1.12) (0.89) (1.17)SOIt4 0.057 0.008 0.041 0.002

(0.60) (0.27) (1.11) (0.06)SOIt5 0.059 0.028 0.087 0.042

(0.65) (0.38) (0.84) (1.15)OOIt1 0.590nnn 0.531nnn 0.402nnn 0.336nnn 0.641nnn

(5.59) (5.12) (4.17) (2.96) (3.57)OOIt2 0.036 0.077 0.088 0.145

(0.27) (0.62) (0.72) (0.59)OOIt3 0.070 0.075 0.090 0.043

(0.71) (0.79) (0.83) (0.30)OOIt4 0.058 0.016 0.094 0.157

(0.49) (0.16) (0.89) (0.94)OOIt5 0.018 0.029 0.088 0.169

(0.16) (0.26) (0.72) (1.06)Rett1 0.003 0.006 0.002

(0.71) (1.21) (0.54)Rett2 0.004 0.006 0.004

(1.15) (1.37) (1.15)Rett3 0.005 0.005 0.005

(1.39) (1.19) (1.49)Rett4 0.001 0.001 0.001

(0.30) (0.37) (0.34)Rett5 0.003 0.000 0.002

(0.62) (0.04) (0.54)OpRett1 1.468 1.581 1.689

(0.87) (0.96) (0.99)OpRett2 1.059 0.517 1.243

(0.64) (0.31) (0.75)OpRett3 0.686 1.816 0.833

(0.45) (1.09) (0.54)OpRett4 1.092 1.098 1.096

(0.80) (0.91) (0.81)OpRett5 0.787 0.528 0.829

(0.55) (0.34) (0.58)Spread 4.258 8.342nn 4.524

(1.32) (1.96) (1.40)Turnover 0.008 0.000 0.011

(0.88) (0.01) (1.23)Volstock 2.773nnn 2.791nnn 2.341nnn

(3.52) (3.18) (3.01)Voloptions 0.763nnn 0.671nn 0.679nnn

(3.41) (2.44) (3.06)No. of obs 1,348,290 1,342,322 1,297,690 1,294,304 892,120 1,294,304Adj. R-squared 0.008 0.009 0.021 0.067 0.063 0.067

J. Hu / Journal of Financial Economics 111 (2014) 625645 635

J. Hu / Journal of Financial Economics 111 (2014) 625645636unambiguous evidence that the options order flow con-tains a significant amount of private information about theunderlying stock0s price movement.

Next, I perform an investment analysis to gauge theeconomic significance of the return predictability. On eachday, the equally weighted quintile portfolios are formedbased on each order imbalance variable. The top quin-

Table 5Excess returns from order imbalance strategies.This table reports the average daily returns on the equally weighted

quintile portfolios based on the total order imbalance (TOI), the stockorder imbalance (SOI), and the options order imbalance (OOI) as well asexcess returns from the longshort portfolios. The portfolios are reba-lanced every day at the market close based on the order imbalancesignals calculated at 3:55 p.m. The excess returns from the longshortportfolios are reported in three forms: the original, the Fama and French(FF) three-factor adjusted, and the four-factor adjusted returns. Allreturns are reported as basis points. Sharpe is the annualized Sharperatio. The resulting t-statistics are reported in parentheses. nnn, nn, and n

denote statistical significance at the 1%, 5%, and 10% level, respectively.

Quintile TOI SOI OOI

Low1 8.720 8.929 2.0952 3.511 2.739 4.4053 0.510 1.804 7.0184 0.856 3.411 4.908High5 11.333 10.396 10.831

51 2.613 1.467 8.736nnn

(1.08) (0.81) (6.03)FF3 alpha 2.552 1.486 8.787nnn

(1.07) (0.85) (6.28)FF4 alpha 2.268 1.228 8.600nnn

(0.95) (0.71) (6.16)Sharpe 0.696 0.426 3.754tile portfolio with the most negative order imbalance isdefined as the sell portfolio, and the bottom quintile port-folio with the most positive order imbalance is definedas the buy portfolio. A zero-investment portfolio is con-structed by buying all of the stocks in the buy portfolio andselling all of the stocks in the sell portfolio at the marketclose. All of the portfolios are rebalanced the next day.Because all of the order imbalance variables are measureddaily at 3:55 p.m., each strategy has five minutes forexecution every day.

Table 5 gives the portfolio returns. The TOI strategygenerates a V-shaped pattern in quintile portfolio returns,and the long-short portfolio does not generate a significantreturn (t-statistic1.08). The average returns on the quin-tile portfolios based on the SOI have the same V-shapedpattern, and no significant return exists for the long-shortportfolio either. The returns on the OOI portfolios increasealmost monotonically across the quintile portfolios. Thesell portfolio has an average return of 2.095 bp per day,and the buy portfolio has an average return of 10.831 bp.The daily excess return reaches 8.736 bp (t-statistic6.03),which amounts to 22% annually. The excess return issignificant at the 1% level after controlling for the Famaand French risk factors and the momentum factor. Theannual Sharpe ratio of the OOI strategy reaches 3.754. Thefindings in this table show that the return predictabilityfrom the OOI is economically significant.5.2. An intraday analysis

To increase the number of observations given thelimited sample length, this subsection investigates thehigh-frequency return predictability from the SOI and theOOI at half-hour intervals. The empirical test is based onthe following equation:

Rett;t1 2

k 1k1 SOItk;tk1

2

k 1k2 OOItk;tk1Ytt1; 6

where Rett;t1 is the stock return calculated using themidpoint of the last NBBO prices in the interval from time tto t1 in which the time unit is half an hour. Yt containsthe NBBO returns for the last two periods, the percentagebidask spread at time t, and the log stock volume and thelog options volume in the last half hour. The firm sub-scription is omitted. The analysis excludes the first halfhour after the market open and the last hour before themarket close and does not use the order imbalances fromthe previous day in the regressions. I choose two lags ofthe imbalance variables in Eq. (6) to balance the goals ofdetecting potential reversal effects and preserving thesample size without missing values. Therefore, the analysisstarts in the fourth half hour of trading (11:00 a.m. to11:30 a.m.) and ends in the 12th half hour (3:00 p.m. to3:30 p.m.) because of the need for the lagged orderimbalance variables. Each stock then has nine half-hourobservations that can be used for estimating Eq. (6) oneach day.

Column 1 of Table 6 gives the Fama and MacBethregression results in the full sample. The SOI positivelypredicts the next half-hour returns with a coefficientestimate of 0.082 (t-statistic2.12). However, the predic-tive relation becomes significantly negative (t-statistic3.86) in the following half hour. TheSOIt2;t1 coefficient is 0.075almost offsetting theimpact of the SOIt1;t completely. The OOIt1;t coefficientis 0.353 (t-statistic3.48), and the second lag OOI has aninsignificant coefficient (t-statistic0.18). The rest of thecolumns show regression results for each half-hour inter-val separately. For example, Column 2 gives the time seriesaverages of the coefficients estimated in the cross-sectional regressions that use only the first observationof the day (11:00 a.m. to 11:30 a.m.). Out of the nine half-hour subsamples, both the SOIt1;t and the SOIt2;t1 havethree coefficients at the 5% significance level, and theOOIt1;t has six coefficients at the 5% significance level.The high frequency results confirm the previous findingsat the daily frequency that the OOI has a permanent priceimpact and that the SOI generates only transitory pricepressure.

5.3. Conditional analysis

Having established the link between the options orderflow and the future stock prices, the analysis now turns toexamining the return predictability from the OOI condi-tioning on option contracts and firm characteristics.

he ordted f

:

midpdent o1 andour;balanles. T

12:30(5)

.789n

2.00).0150.33)0.0681.76)

.314nn

2.01).0910.65).017nn

4.54)0.0010.19)0.8061.39)

J. Hu / Journal of Financial Economics 111 (2014) 625645 637Table 6Predicting stock returns using order imbalances at half-hour intervals.This table investigates intraday relations between the stock returns and t

Fama and MacBeth regression results of the following equation are presen

Rett;t1 2

k 1k1 SOItk;tk1

2

k 1k2 OOItk;tk1Ytt1

The dependent variable Rett,t1 is the stock return calculated by using thethe time unit is half an hour. SOIt1,t is the stock order imbalance independelta imbalance. Yt is a set of control variables at time t, including Rett2,tstock bidask spread; StVolt1,t, the log stock volume in the last half hsubscription is omitted for all variables. The regressions do not use order ima day are excluded from the regression analysis because of missing variabstatistical significance at the 1%, 5%, and 10% level, respectively.

Variable All t11:00 t11:30 t12:00 t(1) (2) (3) (4)

Intercept 2.488nnn 5.249nnn 2.448 4.354nn 2(3.78) (2.77) (1.34) (2.58) (

SOIt1,t 0.082nn 0.093nn 0.117nnn 0.090n 0(2.12) (2.04) (2.69) (1.82) (

SOIt2,t1 0.075nnn 0.106nnn 0.056 0.070nn (3.86) (2.93) (1.50) (2.02) (

OOIt1,t 0.353nnn 0.098 0.540nnn 0.483nnn 0(3.48) (0.47) (3.16) (3.02) (

OOIt2,t1 0.017 0.019 0.067 0.071 0(0.18) (0.14) (0.43) (0.49) (

Rett1,t 0.020nnn 0.023nnn 0.011nnn 0.028nnn 0(14.54) (6.54) (3.00) (7.46) (

Rett2,t1 0.005nnn 0.011nnn 0.002 0.006nn (3.94) (4.34) (0.68) (2.32) (

Spreadt 1.773nnn 2.678nnn 1.017 0.620 (6.82) (3.70) (0.92) (1.04) (5.3.1. Option leverageTable 7 contains the regression results from Eq. (4).

I first show the regression results that use the OOI for eachoption moneyness group separately. Columns 1 and 2show that the OOI constructed by using only the OTMoptions does not predict returns with or without thecontrol variables. This OOI measure is not informative,although the OTM options provide the options investorswith the highest leverage. Columns 36 show that the OOIconstructed by using either the ATM options or the ITMoptions significantly predicts the returns on the next day.The results hold when the OOIs of the three moneynessgroups are examined together in Column 7, and whenthe full control set is added to the model in Column 8.Therefore, the analysis finds no empirical support forHypothesis 3. A possible explanation is that many institu-tional investors use the OTM options to gain volatilityexposures, and the order imbalance does not containmuch information about the underlying stock price.Because the volatility traders usually perform delta hed-ging, their hedging transactions cancel out the optionsmarket makers0 hedging transactions in the stock market.There is little price pressure on the underlying stock either.Another possibility is that the transaction costs associatedwith the OTM options are too high. During the sampleperiod, the average percentage bidask spread for the OTMoptions reaches 11.4%. For the ATM and the ITM options,the percentage bidask spread is only 3.46% and 3%,

StVolt1,t 0.242nnn 0.234 0.190 0.383nnn 0.189(4.12) (1.34) (1.13) (2.73) (1.47)

OpVolt1,t 0.058nnn 0.093nn 0.009 0.114nnn 0.033(4.25) (2.27) (0.16) (3.61) (1.11)er imbalances. Each trading day is divided into 13 half-hour intervals. Theor the full sample as well as for the nine half-hour intervals separately:

oint of National Best Bid and Offer (NBBO) prices from time t to t1, andf option trading. OOIt1,t is the options order imbalance measured as theRett1,t, the NBBO returns for the last two periods; Spread, the percentageand OpVolt1,t, the log options volume in the last half hour. The firmces from the previous day. The first three intervals and the last interval ofhe resulting t-statistics are reported in parentheses. nnn, nn, and n denote

t13:00 t13:30 t14:00 t14:30 t15:00(6) (7) (8) (9) (10)

n 0.396 1.004 3.382nn 1.678 1.885(0.13) (0.69) (2.14) (0.96) (0.88)

0.175nnn 0.062 0.067 0.197 0.318(2.74) (1.45) (0.53) (1.37) (1.02)

n 0.017 0.104nnn 0.003 0.070 0.176(0.32) (2.73) (0.07) (0.77) (1.55)0.446nnn 0.166 0.509nnn 0.271n 0.351nn

(2.79) (0.22) (3.05) (1.72) (2.25)0.168 0.010 0.402 0.309n 0.092(0.89) (0.05) (0.54) (1.94) (0.59)

n 0.034nnn 0.013nnn 0.034nnn 0.006 0.014nnn(7.30) (3.35) (7.64) (1.54) (3.06)0.012n 0.002 0.008n 0.001 0.004(1.92) (0.56) (1.93) (0.13) (0.93)1.546nn 1.801nnn 2.815nnn 1.584nn 3.088nnn(2.00) (2.78) (4.34) (2.35) (4.04)respectively. The spread difference in the option money-ness is consistent with the theoretical prediction ofJohnson and So (2012). If informed traders buy the OTMoptions to gain higher leverage, then they are likely toreverse the trade positions rather than exercise theoptions to take profits. In doing so, they pay the fullamount of the bidask spread, which thereby sharplyincreases their trading costs. The results in Table 7 suggestthat the consideration of the transaction costs can out-weigh the desire for leverage when informed tradersallocate their orders across the options with differentmoneyness.

5.3.2. Level of information asymmetryThis subsection tests Hypothesis 4that the OOI has

greater return predictability for firms with more informa-tion asymmetry. Based on each of the five proxies forinformation asymmetry, I divide the full sample into threegroups: low, medium, and high. I then run full specifica-tion regressions for Eq. (3) in each group. Because themain interest of the test is the predictive ability of the OOIin subgroups, I report in Table 8 only the coefficientestimates and the t-statistics of the OOIt1. Panel A showsthat the OOI coefficient is significant at the 1% level(t-statistic2.85) in the high PIN group, marginally sig-nificant in the medium PIN group (t-statistic1.65), butnot significant in the low PIN group. Panel B shows that theOOI coefficient is significant at the 1% level (t-statistic2.93)

0.078 0.032 0.211 0.466nnn 0.392nn

(0.29) (0.23) (1.46) (3.26) (2.21)0.086nn 0.027 0.027 0.039 0.097nn(2.41) (0.55) (0.79) (1.17) (2.65)

Table 7Return predictability from options order imbalance by option moneyness.This table presents the Fama and MacBeth regression results of the following equation:

Reti;t 5

k 1k1 OTM_OOIi;tk

5

k 1k2 ATM_OOIi;tk

5

k 1k3 ITM_OOIi;tk

5

k 1k4 SOIi;tkXi;t1i;t :

Reti,t is stock i0s return calculated by using the midpoint of the last National Best Bid and Offer (NBBO) prices before the market close on day t. OTM_OOIi,tk, ATM_OOIi,tk, and ITM_OOIi,tk are the options order imbalances calculated by using the out-of-the-money (OTM), at-the-money (ATM), and in-the-money (ITM) option contracts, respectively. SOI is the stock order imbalance unrelated to options. Xi,t1 is a set of control variables on day t1, includingReti,tk, the stock returns for the previous five days; OpReti,tk, the equally weighted options returns across all of the option contracts on stock i for theprevious five days; Spread, the percentage stock bidask spread; Turnover, the ratio of total stock trading volume to number of shares outstanding; Volstock,the log stock volume from 9:45 a.m. to 3:55 p.m.; VolOTM, VolATM, and VolITM, the log trading volumes of the OTM, ATM, and ITM options from 9:45 a.m. to3:55 p.m., respectively. Standard errors are calculated with the Newey and West adjustment to three lags. The resulting t-statistics are reported inparentheses. nnn, nn, and n denote statistical significance at the 1%, 5%, and 10% level, respectively.

Variable (1) (2) (3) (4) (5) (6) (7) (8)

Intercept 5.080 36.136nnn 5.172 31.924nnn 5.172 39.446nnn 5.078 43.958nnn(0.52) (3.01) (0.53) (2.63) (0.53) (3.18) (0.52) (3.46)

OTM_OOIt1 0.578 0.522 0.432 0.388(0.85) (0.89) (0.72) (0.80)

OTM_OOIt2 0.211 0.031 0.136 0.026(0.50) (0.08) (0.34) (0.07)

OTM_OOIt3 0.833 0.046 0.776 0.079(0.77) (0.15) (0.73) (0.25)

OTM_OOIt4 0.727 0.147 0.634 0.145(0.88) (0.46) (0.85) (0.45)

OTM_OOIt5 0.269 0.222 0.178 0.174(0.64) (0.61) (0.44) (0.48)

ATM_OOIt1 0.809nnn 0.646nnn 0.695nnn 0.566nn

(3.00) (2.68) (2.70) (2.33)ATM_OOIt2 0.524 0.465 0.518 0.422

(1.01) (0.95) (1.18) (1.02)ATM_OOIt3 0.417 0.282 0.348 0.332

(1.53) (1.07) (1.37) (1.26)ATM_OOIt4 0.078 0.359 0.166 0.341

(0.2) (1.39) (0.49) (1.33)ATM_OOIt5 0.057 0.120 0.010 0.132

(0.12) (0.48) (0.02) (0.52)ITM_OOIt1 1.084nnn 0.747nn 0.922nn 0.745nn

(2.89) (2.18) (2.56) (2.18)ITM_OOIt2 0.618 0.518 0.572 0.459

(1.61) (1.56) (1.57) (1.27)ITM_OOIt3 0.478 0.279 0.503 0.375

(1.37) (0.83) (1.41) (1.08)ITM_OOIt4 0.525 0.489 0.449 0.374

(1.50) (1.45) (1.47) (1.16)ITM_OOIt5 0.548 0.420 0.529 0.384

(1.51) (1.38) (1.46) (1.30)SOIt1 0.001 0.001 0.002 0.010

(0.02) (0.02) (0.06) (0.25)SOIt2 0.056n 0.051 0.057n 0.056

(1.77) (1.59) (1.75) (1.74)SOIt3 0.038 0.039 0.041 0.038

(1.28) (1.33) (1.39) (1.28)SOIt4 0.006 0.011 0.006 0.005

(0.21) (0.39) (0.22) (0.17)SOIt5 0.029 0.027 0.029 0.032

(0.38) (0.36) (0.40) (0.42)Rett1 0.003 0.003 0.003 0.003

(0.80) (0.77) (0.78) (0.65)Rett2 0.004 0.005 0.004 0.005

(1.14) (1.27) (1.02) (1.24)Rett3 0.005 0.005 0.005 0.005

(1.45) (1.45) (1.39) (1.46)Rett4 0.001 0.001 0.001 0.001

(0.37) (0.26) (0.36) (0.26)Rett5 0.003 0.002 0.003 0.002

(0.64) (0.61) (0.62) (0.59)OpRett1 1.615 1.711 1.772 1.677

(0.95) (1.00) (1.04) (0.98)OpRett2 1.221 1.262 1.379 1.306

(0.74) (0.75) (0.83) (0.79)OpRett3 0.650 0.601 0.595 0.538

J. Hu / Journal of Financial Economics 111 (2014) 625645638

(4)

J. Hu / Journal of Financial Economics 111 (2014) 625645 639Table 7 (continued )

Variable (1) (2) (3)in the low analyst coverage group, significant at the 5% levelin the medium coverage group (t-statistic2.03), but notsignificant in the high coverage group. Panel C shows thatthe OOI has significant coefficients in all of the groups

(0.42) (0.39)OpRett4 1.078 1.156

(0.80) (0.85)OpRett5 0.683 0.666

(0.48) (0.47)Spread 4.453 4.476

(1.37) (1.38)Turnover 0.008 0.008

(0.87) (0.95)Volstock 2.699nnn 2.282nnn

(3.52) (2.99)VolOTM 0.686nnn

(3.39)VolATM 0.545nn

(2.69)VolITM

Table 8Predictive power of options order imbalance and firm characteristics.This table investigates the return predictability from the options order imbalan

sample is divided into three groups: low (o30th percentile), medium (30th70tthe following equation using the Fama and MacBeth regressions:

Rett;t1 2

k 1k1 SOItk;tk1

2

k 1k2 OOItk;tk1Xtt1:

All variables are the same as defined in Table 4. The average slope coefficientsoptions order imbalance (OOIt1). PIN is the probability of informed trading basefollowing. Spread is the percentage stock bidask spread. GH adverse spread is thof Glosten and Harris (1988). LSB adverse spread is the adverse-selection compo(1995). Size is the market capitalization. Ownership is the percentage of shares htraded. Standard errors are calculated with the Newey andWest adjustment to thdenote statistical significance at the 1%, 5%, and 10% level, respectively.

Panel A: PIN

Variable Low Medium High

OOIt1 0.219 0.253n 0.505nnn

(0.74) (1.65) (2.85)

Panel C: Spread

Variable Low Medium High

OOIt1 0.399n 0.429nn 0.901nnn

(1.66) (2.12) (3.03)

Panel E: LSB adverse spread

Variable Low Medium High

OOIt1 0.278 0.513nn 0.516nnn

(1.44) (2.09) (3.28)

Panel G: Ownership

Variable Low Medium High

OOIt1 0.777nnn 0.688nnn 0.237n

(3.64) (2.69) (1.74)(5) (6) (7) (8)based on the bidask spread. However, the coefficient is0.901 (t-statistic3.03) in the high spread group, which ismuch greater than the coefficient of 0.399 (t-statistic1.66) in the low spread group. A similar pattern in the OOI

(0.39) (0.35)1.110 1.236(0.82) (0.92)0.678 0.686(0.48) (0.48)4.512 4.365(1.39) (1.36)0.008 0.009(0.86) (0.96)

2.922nnn 3.394nnn

(3.78) (3.97)0.329nn(2.11)

n 0.158(1.05)

0.904nnn 0.654nnn(4.70) (4.49)

ce for different types of firms. For each of the firm characteristics, the fullh percentile), and high (470th percentile). Within each group, I estimate

and t-statistics (in parentheses) are reported only for the previous day0sd on Easley and O0Hara (1992). Analyst coverage is the number of analystse adverse selection component of the bidask spread based on the modelnent of the bidask spread based on the model of Lin, Sanger, and Bootheld by institutional investors. Options volume is the total options volumeree lags. The resulting t-statistics are reported in parentheses. nnn, nn, and n

Panel B: Analyst coverage

Variable Low Medium High

OOIt1 0.965nnn 0.455nn 0.416(2.93) (2.03) (1.57)

Panel D: GH adverse spread

Variable Low Medium High

OOIt1 0.220 0.358n 0.532nnn

(0.78) (1.87) (2.66)

Panel F: Size

Variable Small Medium Large

OOIt1 1.216nnn 0.276 0.240n

(2.82) (1.24) (1.83)

Panel H: Options volume

Variable Low Medium High

OOIt1 62.376 1.699nnn 0.474nnn

(1.46) (2.66) (4.18)

Table

9Non

linearprice

impactfrom

option

san

dstockorder

imbalances.

Panel

Apresents

theFamaan

dMacBethregression

resultsof

theequation:

Ret

i;t

1SO

I i;t

12SO

I i;t

13OOI i;t

14OOI i;t

1X

i;t1 i

;t;

whereRet

i;tis

stocki0 s

return

calculatedby

usingthemidpointof

thelast

National

BestBid

andOffer

(NBBO)pricesbefore

themarketcloseon

day

t,SO

I i;t

1max

SOI i;t

1;0

,SO

I i;t

1minSOI i;t

1;0

,OOI i;t

1max

OOI i;t

1;0

,an

dOOI i;t

1minOOI i;t

1;0

.Th

econtrol

variable

setXi;t

1isthesameas

defined

inTable4.

Panel

Bpresents

theregression

resultsof

thefollo

wingequation:

Ret

i;t

1SO

I i;t

12OOI i;t

13SO

I2 i;t

1S1 i

;t14OOI2 i;t

1S2 i

;t1X

i;t1 i

;t;

whereS1

i;t1an

dS2

i;t1aredummyvariablesthat

equalon

ewhen

SOI i;t

1an

dOOI i;t

1arepositive,respectively,andnegativeon

eotherwise.Stan

darderrorsarecalculatedwiththeNew

eyan

dWestad

justmen

tto

threelags.T

heresultingt-statistics

arereportedin

paren

theses.n

nn,n

n,andnden

otestatisticalsign

ifican

ceat

the1%

,5%,and10%level,respectively.

Pane

lA:Asymmetricrespon

se

Intercep

tSO

ISO

IOOI

OOI

Ret t1

Ret t2

Ret t3

Ret t4

Ret t5

OpRe

t t1

OpRe

t t2

OpRe

t t3

OpRe

t t4

OpRe

t t5

Spread

Turnover

Vol

stock

Vol

options

34

.384

nnn

0.28

6nnn

0.23

8nnn

0.121

0.80

2nnn

0.003

0.004

0.004

0.002

0.002

0.78

50.43

90.616

0.88

30.52

93.744

0.007

2.612n

nn

0.63

0nnn

(2.86

)(4.24)

(3.63

)(0.71)

(4.63)

(0.71)

(1.09

)(

1.16)

(0.55

)(

0.57

)(

0.45

)(0.27)

(0.40)

(0.65)

(0.38

)(1.17)

(0.76

)(3.40)

(2.93

)

Pane

lB:

Non

linearity

Intercep

tSO

IOOI

S 1SOI2

S 2O

OI2

Ret t1

Ret t2

Ret t3

Ret t4

Ret t5

OpRe

t t1

OpRe

t t2

OpRe

t t3

OpRe

t t4

OpRe

t t5

Spread

Turnover

Vol

stock

Vol

options

33.537

nnn

0.12

2nn

1.111n

nn

0.001

0.02

0nnn

0.003

0.004

0.004

0.002

0.002

0.87

90.38

00.43

61.108

0.48

84.23

30.008

2.66

1nnn

0.75

8nnn

(2.77

)(2.53)

(5.75)

(1.39

)(

3.50

)(0.62)

(1.24

)(

1.20

)(

0.56

)(

0.55

)(

0.50

)(0.23)

(0.28)

(0.81)

(0.35

)(1.32)

(0.87)

(3.43)

(3.36

)

J. Hu / Journal of Financial Economics 111 (2014) 625645640coefficients is found in groups based on the adverse-selection component of the spread in Panels D and E,and in the groups based on the firm size in Panel F.Collectively, these results suggest that, consistent withHypothesis 4, the OOI has a stronger predictive powerfor returns when the firm has a higher level of informationasymmetry.

5.3.3. Institutional ownershipThis subsection describes the testing of Hypothesis

5that the information content of the OOI varies witha stock0s institutional ownership. Panel G in Table 8 givesthe results. The OOI is statistically significant at the 10%level in all of the ownership groups. However, the OOIcoefficient is 0.777 (t-statistic3.64) in the low ownershipgroup and 0.237 (t-statistic1.74) in the high ownershipgroup. The differences in the magnitude of the coefficientsand the t-statistics show that the OOI has a larger amountof price information for the firms with low institutionalownership. These stocks are associated with high short-sale costs. The results suggest that short-sale constraints inthe stock market increase the information content in theoptions order flow.

5.3.4. Market activityHypothesis 6 is tested in this subsection. Panel G in

Table 8 shows that the OOI does not significantly predictthe underlying returns when the options trading volume islow. The OOI coefficients are statistically significant at the1% level in both the medium (t-statistic2.66) and thehigh volume groups (t-statistic4.18). This finding sug-gests that the predictive ability of the OOI comes from theperiods when the options market is active. The results areconsistent with the information explanation for the returnpredictability of the OOI.

6. Further analysis

Section 5 shows that the options order flow containsimportant price information about the underlying stock.In this section, I perform additional analyses to better under-stand that information linkage.

6.1. Asymmetric price response

I first investigate whether there is asymmetric priceresponse to the imbalances by estimating the followingequation:

Reti;t 1 SOIi;t12 SOIi;t13 OOIi;t14 OOIi;t1Xi;t1i;t ; 7

where SOIi;t1 maxSOIi;t1;0, SOIi;t1 minSOIi;t1;0,OOIi;t1 maxOOIi;t1;0, and OOIi;t1 minOOIi;t1;0.The results are reported in Panel A of Table 9. The 1estimate is 0.286, statistically significant at the 1% level (t-statistic4.24). The 2 estimate is 0.238, also significantat the 1% level (t-statistic3.63). These results suggestthat both the positive and negative SOIs significantly predictstock returns, but the signs are inconsistent. Therefore, theoverall SOI does not exhibit an ability to predict stock

conveys information about corporate earnings five days

J. Hu / Journal of Financial Economics 111 (2014) 625645 641returns. The 3 estimate is insignificant, and the 4 estimateis significant at the 1% level (t-statistic4.63). These resultssuggest that the predictive power of the OOI mainly comesfrom the negative OOI, which support Hypothesis 5 thatinformed traders use options to get around the short-saleconstraints in the stock market when they acquire negativeinformation.

6.2. Nonlinear price impact

This subsection investigates the nonlinearity in thepredictive relation. Both the SOI and the OOI have fat tails.The possibility exists that the stock price0s response isdifferent in the tails of the order imbalances. I estimate thefollowing specification:

Reti;t 1 SOIi;t12 OOIi;t13 SOI2i;t1S1i;t14 OOI2i;t1 S2i;t1Xi;t1i;t ; 8

where S1 and S2 are dummies that equal one when SOIand OOI are positive, respectively, and negative oneotherwise. The interaction terms quadratically amplifythe two imbalance variables and parse out the tail effect.Panel B of Table 9 gives the regression results. Both the SOIand the OOI coefficients are positive and statisticallysignificant at the 5% level. No significant nonlinear effectfor the SOI is found because the 3 is insignificant.However, the 4 is negative and significant at the 1% level(t-statistic3.5), indicating that the predictive ability ofthe OOI reduces in the tails.

6.3. Moving average and shocks

Both the SOI and the OOI have positive autocorrela-tions. In this subsection, I decompose each imbalancevariable into a moving average (MA) component and ashock component to further investigate the sources of theprice impact. For example,

OOIOOIMAk OOIshockk ; 9where the OOIMAk is the average OOI from the previous ktrading days. For simplicity, I omit the firm and timesubscriptions in Eq. (9). The same decomposition is alsoperformed for the SOI. Then, I estimate the followingequation:

Reti;t 1 SOIMAi;t 1 2 SOIshocki;t 1 3 OOIMAi;t 14 OOIshocki;t 1 Xi;t1i;t ; 10

Three MA periods of three days, five days, and ten days arechosen, and the results are reported in Table 10. Panel Ashows that the correlations between the MAs and theshocks are large and negative for both the SOI and the OOI.For example, the three-day correlations are 0.468 and0.514 for the SOI and the OOI, respectively. As the lengthof the MA increases, the correlations become weaker.

Panel B reports the regression results from Eq. (10).The first column reports the result from using a three-dayMA decomposition. Although the MAs and the shocks arenegatively correlated, both are positively correlated withthe future stock returns. Neither the MA nor the shock ofthe SOI significantly predicts the stock returns. For the OOI,before the announcement.Informed trading should be more active when the

potential profit is high. The last empirical test investigateswhether the OOI also becomes more informative whenapproaching such earnings annou