Forecasting digital interactions
Transcript of Forecasting digital interactions
![Page 1: Forecasting digital interactions](https://reader030.fdocuments.us/reader030/viewer/2022021010/6204427507fdc37291327e37/html5/thumbnails/1.jpg)
Forecasting digital interactionsPhilip Stubbs, Atlantic Insight Thursday 14 October 2021
![Page 2: Forecasting digital interactions](https://reader030.fdocuments.us/reader030/viewer/2022021010/6204427507fdc37291327e37/html5/thumbnails/2.jpg)
The importance of data accuracy
#1 Make sure your historical data is accurate
#2 Understand in detail the sources of digital volume data
#3 Automate the assembly of digital volumes
#4 Understand when demand was constrained
Chat graph
#5 Data Preparation – fix data before passing it into a model
![Page 3: Forecasting digital interactions](https://reader030.fdocuments.us/reader030/viewer/2022021010/6204427507fdc37291327e37/html5/thumbnails/3.jpg)
Understand the historical variation in digital volumes
#6 Understand the variation in historical volumes
#7 Chart historical data and label with causes of historical variation
#8 Keep a diary of reasons for variation in digital volumes
#9 Use terms Trend and Seasonality correctly
Xmas campaigns
1st lockdown
Xmas camps & 2nd lockdown
Marketing activity
Marketing activity
Trend is a long term increase or decrease in
the data
Seasonality is where a time series is
affected by seasonal factors such as time of
year or day of the week
![Page 4: Forecasting digital interactions](https://reader030.fdocuments.us/reader030/viewer/2022021010/6204427507fdc37291327e37/html5/thumbnails/4.jpg)
Understand future impacts to digital volumes
#10 Identify individuals whose actions contribute to variation
#11 Get accurate information about digital volume drivers
with enough time to change the resource plan
#12 Challenge the accuracy and timeliness of drivers
#14 Be conservative with respect to promises of a new system and/or process
#13 Keep very close to website changes & changes to customer communications
![Page 5: Forecasting digital interactions](https://reader030.fdocuments.us/reader030/viewer/2022021010/6204427507fdc37291327e37/html5/thumbnails/5.jpg)
Forecasting process
#17 Graph forecasts next to recent actuals
#15 Judgement from stakeholders can improve accuracy
#16 Study errors & share with clients and suppliers
Emails received
![Page 6: Forecasting digital interactions](https://reader030.fdocuments.us/reader030/viewer/2022021010/6204427507fdc37291327e37/html5/thumbnails/6.jpg)
Other forecasting points
#18 Channel shift will change the total number of enquiries
#19 Using scenarios more important when uncertainty greater
#20 Automate forecasting when conditions are right
• Automated forecasts are more accurate analysis + judgement
• Errors on these digital channels don’t have CX / commercial impact
• Forecast analyst support is limited
![Page 7: Forecasting digital interactions](https://reader030.fdocuments.us/reader030/viewer/2022021010/6204427507fdc37291327e37/html5/thumbnails/7.jpg)
Time series models
#21 Understand Time Series models
Only look at historical values of what we are trying to forecast – no external variable involved
• Simple averages
• Weighted averages
• Bespoke model
• Exponential Smoothing, eg Holt Winters
• ARIMA
![Page 8: Forecasting digital interactions](https://reader030.fdocuments.us/reader030/viewer/2022021010/6204427507fdc37291327e37/html5/thumbnails/8.jpg)
Causal models
#23 Make sure three assumptions are in place before you use a
Causal model:
#24 Time Series models often outperform Causal models
#22 Try regression #25 Dig a little deeper to find a causal model
1. A strong relationship exists with the driver
2. Expect the relationship will continue to exist in the future
3. Reliable forecasts of the driver exist
Produces a y=mx + c
model that can be used as
an equation to estimate
future emails
Nu
mb
er o
f e
mai
ls p
er
we
ek
Initial Regression Model
This Weeks Emails = 834 + 0.134 * This Weeks Orders
Lagged Regression Model
This Weeks Emails = 892 +
0.079 * Last Weeks Orders +
0.053 * Week -3 Orders
![Page 9: Forecasting digital interactions](https://reader030.fdocuments.us/reader030/viewer/2022021010/6204427507fdc37291327e37/html5/thumbnails/9.jpg)
Final points
#27 Try Machine Learning models
#28 Look into SAS, R or Python
#29 Further Reading (buy the latest editions)
#26 Select model based on its forecasting ability, and not on model fit
![Page 10: Forecasting digital interactions](https://reader030.fdocuments.us/reader030/viewer/2022021010/6204427507fdc37291327e37/html5/thumbnails/10.jpg)
Forecasting Digital InteractionsPhilip Stubbs, Atlantic Insight Thursday 14 October 2021