Reducing Volatility in Multi-Horizon Forecasting
Forking-Sequences brings stability to forecast revisions.
1What is Forking-Sequences?
Forking-Sequences is a neural network architecture that jointly encodes and decodes time series forecasts across all forecast creation dates (FCDs) in a single forward pass. This approach offers a statistically and computationally more efficient training paradigm compared to traditional window sampling methods, providing significant advantages especially in multi-horizon forecasting scenarios.
2Why Forecast Volatility Matters
While accuracy is crucial in forecasting models, it's not the only metric that counts in production environments. Forecast volatility refers to the consistency of prediction revisions across different FCDs for the same target date. Excessive volatility can undermine trust and complicate planning for users like electrical grid operators.
3New Volatility Metrics
The research introduces two novel volatility metrics: scaled Forecast Percentage Change (sFPC) measures real-time revision magnitude, while Excess Volatility (EV) specifically penalizes revisions that move forecasts away from truth or overshoot. These metrics help users distinguish between beneficial and harmful revisions.
4Volatility Reduction with Exponential Smoothing
Exponential smoothing forecast ensembling (α = 0.9) reduces sEV by 10-13% across all encoder types. Most remarkably, this improvement comes with less than 0.1% accuracy degradation. The method works out-of-the-box even with models pretrained using window sampling.
5Testing on Time Series Foundation Models
The method was tested on Time Series Foundation Models (TSFMs) including Chronos-2, Toto 2.0, TimesFM, PatchTST, and N-BEATS. Results show approximately 10% reduction in volatility with negligible accuracy cost, demonstrating the approach's broad applicability across different model architectures.
6Electrical Grid Example
When an electrical grid operator revises load forecast from 45 GW to 65 GW ahead of a heat wave, it's a useful revision. However, erratic jumps between FCDs without new information complicate operational planning. Forking-Sequences effectively reduces such excessive revisions while maintaining forecast quality.
7Computational Efficiency
By reusing encoder computations, Forking-Sequences makes forecast ensembling computationally efficient. This presents a significant advantage over window sampling methods. The system can produce multiple forecasts without requiring additional encoder computations.
8Industrial Applications
The method has potential applications in energy management, logistics planning, and financial forecasting. Particularly in scenarios requiring real-time decision making, forecast stability can enhance operational efficiency. Developed by Amazon's SCOT team, the method has been successfully tested in internal operations.
9Academic Contribution
Published in Transactions on Machine Learning Research in 2026, this collaborative research between Carnegie Mellon University and Amazon has the potential to set new standards in time series forecasting. The paper is available on arXiv with DOI: 10.48550/arXiv.2510.04487.
10Future Work
Researchers plan to further test the method's performance across different industrial scenarios. Particular focus will be on applications in climate modeling and healthcare predictions. Additional work continues to refine the volatility metrics for broader applications.
11Conclusions and Implications
Forking-Sequences and forecast ensembling represent a breakthrough in time series forecasting. By maintaining accuracy while reducing volatility, it promises reliable predictions for industrial applications. Widespread adoption could improve operational decision-making in sectors like energy, logistics, and finance.
Original article: ML@CMU — Çoklu Ufuklu Tahminlerde Volatilite Azaltma Yöntemi