Experts of the Industrial Organization and Infrastructure Economics Department at the Gaidar Institute discussed new specialized architectures in time series forecasting: NBEATS and the recently introduced NHITS. These are replacing classic recurrent neural networks (such as LSTM and GRU).
NHITS (Neural Hierarchical Interpolation for Time Series) is a new deep model specifically adapted for time series forecasting. Its key feature is the combination of the advantages of ensemble approaches (such as random forest), where complex patterns are described using a set of simple models and recurrent neural networks, where forecasts are built recursively based on previous observations.
Experts highlighted main NHITS principles:
- The model recursively decomposes the forecasting task into several levels of detail.
- At each level, simple local models explain the portion of the time series not described in previous level.
Experts identified two main advantages of NHITS over classical recurrent models. First, the model is more robust to noise and complex patterns due to its multi-level explanation. Second, the hierarchical approach reduces the risk of overfitting by avoiding parameter overload, which is typical for many recurrent networks.
The Industrial Organization and Infrastructure Economics Department use NHITS for forecasting the index of industrial production, where compared to the ARIMAX benchmark, the neural network managed to improve the Directional Accuracy metric (the percentage of correctly predicted direction of change) on test data from 60% to 75% with statistically indistinguishable mean squared forecast errors. This result is explained by two main factors:
- ARIMA does not find sufficiently stable linear dependencies and its forecast degenerates into a forecast of the previous value (a naive forecast), which, although optimal from the point of view of minimizing the expected squared error, is poorly suited for the tasks of forecasting the direction of change in an indicator.
- The NHITS model, due to its nonlinear architecture and hierarchical structure, was able to identify dependencies in multivariate data. This overcame the naive forecasting limitation caused by excessive reliance on the most recent observed value of the series.