Deep Learning for Sales Forecasting and Lift Prediction

Deep Learning for Sales Forecasting and Lift Prediction

CHALLENGE

Sales forecasting is an essential tool for business management, enabling companies to make informed business decisions for both short-term and long-term planning. By accurately predicting achievable sales revenue,  critical insight is derived which will guide resource allocation in order to plan for anticipated growth and avoid unforeseen cash flow problems.

Lift knowledge is also essential in order to understand how customers will convert to a company’s products in response to marketing and advertising efforts. Consequently, an ability to forecast sales lift will assist with optimized planning of marketing campaigns.

However, sales data often incorporates high variability, and the predictive relationships might be too complex to accurately capture by traditional modeling methods. This poses a significant challenge, because if the model cannot fully learn the patterns within the data set, then it cannot be relied upon to make accurate predictions.


SOLUTION

AlgoTactica has directed significant R&D effort towards identifying reliable modeling procedures for very complex data sets. In these cases, interaction patterns between variables are not static, but evolve according to the magnitude of the variables, in addition to other factors. This requires that the model be able to dynamically alter its forecast strategy depending on the input values it is given for each predictive action; traditional linear modelling techniques cannot achieve this.

In our experience, deep learning neural networks offer the best potential for these situations, and have the highest degree of predictive accuracy. Here, we briefly discuss results obtained by using deep learning to predict sales of a snack item for a well-known grocery chain.


RESULTS

In this case study, linear regression and deep learning models were designed to forecast in-store revenues for several brands of a dry snack item, based on price level and promotional strategy, as well as other predictors. The Linear Regression: Snack Sales scatterplot reveals that when actual sales are compared to predicted sales, the traditional linear model performs very poorly, and offers no predictive power at all, with an r-squared of only 0.53. However, the Deep Learning: Snack Sales scatterplot indicates that a 3-layer deep neural net does offer useful forecasting power, with an r-squared of 0.85. While this was an extremely complex data set, the deep learning model has learned the data patterns sufficiently well to enable sales trending forecasts. There is still some error, but investigation of outliers for possible inclusion in a separate extremal model might lead to further improvements.

Once a deep learning model has been adequately fitted, it can then be used to simulate possible outcomes, for a range of inputs. The Predicted Sales By Pricing Strategy graph illustrates predicted sales for this product at a specific store, over a range of pricing levels in combination with sales promotional strategies. By using the adapted model in this simulation mode, it is possible to estimate the lift associated with various marketing scenarios, and use this information as guidance when planning sales promotions. For the case shown here, further investigation of customer behavior is also warranted, because the analysis reveals an unusual sales demand curve in response to pricing levels during periods of non-promotion.

-->
Back to top