Pastprop-RNN: improved predictions of the future by correcting the past

TL;DR

Pastprop-RNN introduces a data-centric backpropagation method called Pastprop-LSTM that improves forecasting accuracy by adjusting training data to better explain future outcomes, especially in noisy or anomalous datasets.

Contribution

The paper presents Pastprop-LSTM, a novel backpropagation algorithm that assigns responsibility to training data, enhancing forecasting models' robustness and accuracy.

Findings

Improves forecasting accuracy on M4, M5, and Numenta datasets.

Effective in datasets with anomalies and high error rates.

Demonstrates potential for data correction in time series prediction.

Abstract

Forecasting accuracy is reliant on the quality of available past data. Data disruptions can adversely affect the quality of the generated model (e.g. unexpected events such as out-of-stock products when forecasting demand). We address this problem by pastcasting: predicting how data should have been in the past to explain the future better. We propose Pastprop-LSTM, a data-centric backpropagation algorithm that assigns part of the responsibility for errors to the training data and changes it accordingly. We test three variants of Pastprop-LSTM on forecasting competition datasets, M4 and M5, plus the Numenta Anomaly Benchmark. Empirical evaluation indicates that the proposed method can improve forecasting accuracy, especially when the prediction errors of standard LSTM are high. It also demonstrates the potential of the algorithm on datasets containing anomalies.