Sign in or start a free trial to avail of this feature.
9. Running a Covariate Forecast
We continue our examination of covariate forecasting techniques in this lesson. In particular, we will consider how a change in budgeted marketing spend might influence future sales.
TS Covariate Forecast
- This tool provides forecasts for a time series model with covariates
- One input for this tool must be contain the data stream with the ARIMA or ETS model
- The other input must contain future covariate information
In the previous lesson, we learned that our marketing budget will increase substantially in the year ahead. Given this information, we'd like to know how it might affect our online sales forecast. To prepare for our forecast, we aggregated the historic online marketing cost into a single field. We then consolidated this data to the weekly level and merged it with the weekly marketing budget for 2017.
In this lesson, we'll pick up where we left off and run our forecast on the combined 2017 budget and historic data.
We'll accomplish this goal by following four key steps. First, we'll run our historic data through an ARIMA model as a base for our covariate forecast. We'll then run our covariate forecast by connecting both our ARIMA model and the full data set to the TS Covariate Forecast Tool. Third, we'll create a univariate ARIMA forecast so we can determine the effect of the increased marketing budget. Finally, we'll join and summarize the two forecasts and analyze the results.
We'll start this lesson by running our historic data through an ARIMA model. In order to run the TS Covariate Forecast Tool, we'll need to obtain the output for the original ARIMA forecast. However, as we now have a range of dates, we must begin by specifying the date range for the historic data. To that end, we'll navigate to the Preparation tab on the Tools palette and connect a date filter tool to our workflow.
In the Configuration window, we'll select a starting date of January 6th, 2014 and an ending date of 26th December, 2016.
These dates represent the start and end of our historical weekly data. we'll now connect an ARIMA tool to the date filter and use the same configuration settings from the previous lessons. In the Other Options tab, we'll set the starting year as 2014 and the number of periods as 52.
In the Required Parameters tab, we'll target weekly sales with a field frequency of weekly. On this occasion, we'll select the use covariates box and choose the weekly online marketing field in the options presented.
We'll then name this model ARIMA_covariate.
At this point, we can move on to step two and run our covariate forecast. We'll start by bringing the TS Covariate Forecast Tool onto the canvas. We'll connect the output from the ARIMA model to the left input node of the covariate tool. The other input mode must join to the dataset but include the period with the budget information for the weekly online marketing spend. As such, we'll bring another date filter tool onto the canvas, connecting it to both the union tool and the right input node of the TS Covariate Forecast Tool.
Again, we'll set the start date as January 6th, 2014, but this time, set the end date of December 25th, 2017.
To ensure that we interpret the output from the covariate tool correctly, we'll navigate back to the Configuration window for the covariate tool and change the field name to Covariate_forecast.
We'll now run the workflow. Note that processing this calculation could take some time, so I'll skip ahead.
If we view the output from the covariate tool, we can see that it specifies the 52 weeks of the budget.
We'll now move on to step three and create a univariate ARIMA forecast for comparison. In order to understand how this budget changes the original forecast, we'll want to align the two data sets up together. We can do this by duplicating the existing ARIMA tool in a parallel workflow. We'll right-click the existing ARIMA tool and click Copy.
We'll then paste the ARIMA tool to another part of the canvas, reconnect it to the historic dates, uncheck the use covariates box and rename it ARIMA_univariate.
We'll bring down a TS forecast tool and connect it to the output from the ARIMA tool.
In the Configuration window of the forecast tool, we'll set the number of periods to 52 and rename the field Original_forecast so we can easily distinguish outputs from the two models. For the final step of this lesson, we'll join and summarize the two forecasts and analyze the results.
This will allow us to understand the effect of the increased marketing budget. We'll now bring down a join tool and connect the two forecasts. In the configuration window of the join tool, we'll link the forecasts by period and subperiod.
We'll connect a browse to the J output node and run the workflow.
Again, this could take some time so I'll skip ahead.
We now have both forecasts presented together, but what is this information telling us? From here, we can connect a summarize tool to compare the aggregate forecast results. We'll sum the sales for the original forecast, as well as for the covariate forecast.
We'll run the workflow again to view our final results.
As we can see, the original forecast projected sales of roughly 1.9 million, while the covariate forecast projected sales of just over 2.1 million.
The difference between these forecasts is just 200,000.
Recall that we're increasing our marketing spend from 0.89 million to 1.2 million.
This is an increase of 310,000.
In other words, we're spending an extra 310,000 to earn a marginal revenue of just 200,000. Even if our incremental margin on this extra revenue is 100%, there are very few scenarios where such an expense makes financial sense.
Before ending the lesson, let's quickly recap what we've done. First, we ran our historic data through an ARIMA model as a base for our covariate forecast. We then ran our covariate forecast by connecting both our ARIMA model and the full data set to the TS Covariate Forecast Tool. Next, we created a univariate ARIMA forecast so we can compare it to our covariate forecast. Finally, we joined and summarized the two forecasts and analyzed the results. In the next lesson, we'll take a brief look at why you might want to adjust your ETS or ARIMA model settings and what effect it could have on your outputs.