How Far Can You Actually Forecast? The Lyapunov Ceiling

2026-04-17 · Source: Valeriy’s Substack · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, quick

Summary

The provided content describes a common issue in time series forecasting where model performance degrades significantly as the prediction horizon increases. Initial one-step-ahead forecasts may appear accurate, but extending the forecast to 5, 10, or 20 steps rapidly leads to a decline in quality, often resulting in "garbage" predictions. This degradation rate appears consistent, suggesting that simply adding more network depth, features, or data does not inherently resolve the problem of decaying forecast quality over longer horizons.

Key takeaway

For Research Scientists developing time series models, recognize that forecast quality inherently decays with longer prediction horizons. Instead of solely focusing on model complexity or data volume, investigate methods specifically designed to maintain accuracy over extended periods, or consider ensemble approaches that combine short-term and long-term forecasting strategies to mitigate rapid degradation.

Key insights

Forecast quality in time series models decays predictably with increasing prediction horizon.

Principles

• Longer horizons degrade forecast quality
• More data does not prevent decay

Topics

• Forecast Quality
• Prediction Horizon
• Lyapunov Ceiling
• Model Performance
• Forecasting Limitations

Best for: Research Scientist, Machine Learning Engineer, Data Scientist, AI Scientist

Related on AIssential

• The Sigmoids Won't Save You — Exponential trends often persist longer than predicted, defying premature sigmoid curve assumptions.
• Is Your Time Series Even Forecastable? — Forecastability is a data property, not a model property, determining a time series' theoretical prediction ceiling.
• Bridging the Last Mile of Time Series Forecasting with LLM Agents — LLM agents can bridge the gap between statistical time series forecasts and business-ready predictions by integrating weakly structured contextual information.
• Forecast collapse of transformer-based models under squared loss in financial time series — Highly expressive models like Transformers degrade on financial time series due to increased variance from noise reuse.
• Benchmarking Counterfactual Prediction in Epidemic Time Series with Time-Varying Interventions — A new benchmark provides realistic counterfactuals for epidemic time series, revealing challenges in causal inference methods.
• ​​Time Series Cross-Validation: A Guide to Techniques & Practical Implementation — Time series cross-validation maintains chronological order for robust model evaluation and accurate forecasting.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Valeriy’s Substack.