• Vinod Joshi
• R.K. Somani

Time series forecasting is generally used for making informed decisions in fields such as energy, finance, and operations, yet traditional models like ARMA often fail to forecast recent data. This research addresses the challenge by proposing an Adaptive Decay-Weighted ARMA model, which introduces a learnable decay-weighted loss function to dynamically operate on recent observations and reduce the importance of older data. The model combines different exponential decay functions, moving averages, and seasonal feature tuning. Therefore, our model provides flexibility to adapt to diverse temporal patterns. Also, validation on real-world datasets, i.e., U.S. electricity production, demonstrates that the proposed approach consistently outperforms standard models such as Normal AR, ARMA (1,1), and AR with cycle-only features. The results are impressive when the proposed model achieved a Mean Absolute Percentage Error (MAPE) as low as 1.04\% for short-term forecasts, with higher accuracy in the multiple forecasting benchmarks. These results confirm that adaptive weighting significantly enhances predictive performance. Hence, our model contributes to recent advancements in weighted ARMA models and adaptive hybrid forecasting methods. The proposed approach requires additional computational resources for training, but it can be integrated with neural networks or attention mechanisms, which makes it a robust solution for practical forecasting. This research's future work will explore advanced decay learning strategies and online adaptation to further improve real-time forecasting capabilities.

Time Series Forecasting, Adaptive Decay-Weighted ARMA

"Dynamic Recency Weighting in ARMA Models: Bridging Classical and Deep Learning Approaches for Time Series Forecasting", International Journal of Emerging Technologies and Innovative Research (www.jetir.org), ISSN:2349-5162, Vol.12, Issue 5, page no.l361-l370, May-2025, Available :http://www.jetir.org/papers/JETIR2505C29.pdf

"Dynamic Recency Weighting in ARMA Models: Bridging Classical and Deep Learning Approaches for Time Series Forecasting", International Journal of Emerging Technologies and Innovative Research (www.jetir.org | UGC and issn Approved), ISSN:2349-5162, Vol.12, Issue 5, page no. ppl361-l370, May-2025, Available at : http://www.jetir.org/papers/JETIR2505C29.pdf