Interpretable Sequence Models for the Sales Forecasting Task: A Review

Sequence modelling has shown tremendous potential in solving real-world sequence prediction tasks like speech recognition, time series forecasting, and context identification. However, most of these sequence models are trained on univariate datasets and cannot leverage the information available in a multivariate setting. Moreover, the prediction/decision made by these models is not interpretable; consequently, the end users are unaware of the different steps involved in reaching that prediction/decision and cannot determine if the model aligns with the business and ethical values. This work investigates the performance of different sequence learners trained in a multivariate setting for the sales forecasting task. Specifically, different sequence models, including vanilla LSTM, stacked LSTM, bidirectional LSTM, and convolution neural networkbased-LSTM, have been trained on the Walmart dataset, and a comparative analysis of their performance using mean squared error (MSE) and weighted mean absolute error (WMAE) metric is reported. For training the learners in a multivariate setting, relevant features have been identified using exploratory data analytics. Furthermore, these sequence models are made interpretable using the Local Interpretable Model Agnostic Explanation (LIME) model to explain away the key variables involved in the prediction task. Empirical results obtained on the Walmart sales dataset established that the performance of the stacked LSTM model is superior to other learners. Additionally, the stacked model being the most generalizable, is complemented by the LIME module to explain away its predictions using the relevant features.