Author(s): Shenggang Li

Originally published on Towards AI.

Integrating Recency-Based Loss Weighting and Seasonal Feature Tuning for Enhanced Predictive AccuracyPhoto by Haberdoedas II on Unsplash

Time series forecasting is both fascinating and challenging. It’s fascinating because accurate predictions can directly inform better decisions — whether it’s managing electricity demand, planning inventory, or making investment moves. The ability to anticipate future values based on past patterns is a powerful tool in many fields. But forecasting is also hard, mainly because time can shift trends, disrupt cycles, and introduce new behaviors that make modeling complex.

Traditional models like AR(k) or ARMA have long been used for this task. However, one key limitation is that they treat all past observations equally, assuming each lag contributes the same to predicting the future. In practice, though, recent values are often more relevant than older ones — something these classical models tend to overlook.

I now proposed a new method called Adaptive Decay-Weighted ARMA. The idea is simple but powerful: I adjusted the usual loss function by adding a decay weighting function, F(sequence_j∣A), which assigns higher importance to recent observations and less to distant ones. The parameter A controls how fast this decay happens. For example, the most recent value (sequence_1) always gets full weight, and older values fade depending on the decay pattern. What’s more, I don’t… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI