Optimizing Sales Forecasts through Automated Integration of Market Indicators

TL;DR

This work integrates macroeconomic time series from the Eurostat database into Neural Prophet and SARIMAX forecasting models and finds the Forward Feature Selection technique consistently yielded superior forecasting accuracy for both SARIMAX and Neural Prophet across different company sales datasets.

Abstract

Recognizing that traditional forecasting models often rely solely on historical demand, this work investigates the potential of data-driven techniques to automatically select and integrate market indicators for improving customer demand predictions. By adopting an exploratory methodology, we integrate macroeconomic time series, such as national GDP growth, from the \textit{Eurostat} database into \textit{Neural Prophet} and \textit{SARIMAX} forecasting models. Suitable time series are automatically identified through different state-of-the-art feature selection methods and applied to sales data from our industrial partner. It could be shown that forecasts can be significantly enhanced by incorporating external information. Notably, the potential of feature selection methods stands out, especially due to their capability for automation without expert knowledge and manual selection effort. In particular, the Forward Feature Selection technique consistently yielded superior forecasting accuracy for both SARIMAX and Neural Prophet across different company sales datasets. In the comparative analysis of the errors of the selected forecasting models, namely Neural Prophet and SARIMAX, it is observed that neither model demonstrates a significant superiority over the other.

Figures (7)

• Figure 1: This figure shows forecasts of SARIMAX without and with one exogenous variables. Both models were trained on the normalized BC1_64 data. One with adding an exogenous variable from Eurostat dataset. This plot only shows the time range from 2020 on to have better visibility on the out-of-sample behavior of the forecasts (2021.05-2022.04). The values for the exogenous variable (in black) are only shown for the training data time range (until 2021.04) as these are the values taking into account from the model. All plots in this work are self-generated with the Python library matplotlib or MS PowerPoint.
• Figure 2: This figure shows an overview of the procedure applied by the implementation of the FFS in this work. OOS MAE stands for out-of-sample mean absolute error.
• Figure 3: This figure shows the NP model (left) and SARIMAX model performing on the different time ranges of normalized training data (BC1_64 (2016) and BC1_28 (2019)) without adding exogenous variables. This plot only shows the time range from 2020 on to have better visibility on the out-of-sample behavior of the forecasts (2021.05-2022.04).
• Figure 4: This figure shows NP model's forecast when adding the exogenous variables found with the Forward Feature selection using NP model vs. the forecast of the model without exogenous variables for BC3_64. This plot only shows normalized data for the time range from 2021 on to have better visibility on the out-of-sample behavior of the forecasts (2021.05-2022.04).
• Figure 5: This figure shows the mean OOS MAE score development per number of selected variables for the Forward Feature Selection for all the Forward Feature Selection experiments discussed in this work.