tofee-tree: automatic feature engineering framework for modeling trend-cycle in time series forecasting

Most time series forecasting tasks using Artificial Neural Networks (ANNs) relegate trend-cycle modeling to a simple preprocessing step. In this work, we propose an automatic feature engineering framework for modeling the trend-cycle (tofee-tree) in time series forecasting. The first stage of the framework automatically creates over 286 deterministic linear and nonlinear engineered features to model the trend-cycle. These features are based only on the time of observation and length of the time series, making them domain-agnostic. In the second stage of the framework, a SHapley Additive exPlanations (SHAP)—based feature selection procedure using Light Gradient Boosted Machine (LightGBM) selects the most relevant features. These relevant features can be used for forecasting with ANNs in addition to the auto-regressive lags. Two popular ANNs—Multi-Layer Perceptron (MLP) and Long Short Term Memory network (LSTM) are used to evaluate our proposed tofee-tree framework. Comparisons against two empirical studies using the M3 competition dataset show that the proposed framework improved the overall Symmetric Mean Absolute Percentage Error (SMAPE) in the one-step, medium- and long-term. The relative improvement in one-step SMAPE is 3% for MLP and 23% for LSTM. We also show that the residual seasonality left after deseasonalization can be modeled using the tofee-tree framework.