Leaky-integrator reconstruction: taming error accumulation in recursive differenced time-series forecasting

TL;DR

The paper introduces leaky-integrator reconstruction, a method to reduce error accumulation in recursive differenced time-series forecasting. This approach provably bounds the accumulated error variance by moving the pole inside the unit circle. It can be applied to any deployed forecaster with a simple two-line change.

Why It Matters

The proposed method has significant practical implications as it can improve the accuracy of time-series forecasting models, which are crucial in various fields such as finance and economics. By reducing error accumulation, leaky-integrator reconstruction can lead to more reliable predictions and better decision-making. This can have a substantial impact on industries that rely heavily on forecasting, such as stock markets and supply chain management.

Key Findings

• Recursive differenced forecasting can lead to error accumulation and divergence
• Leaky-integrator reconstruction can bound the accumulated error variance
• The method can be applied to any deployed forecaster with a simple modification