The Unreasonable Difficulty of Time Series Forecasting

Preview

Preview I’ve been thinking recently about what makes time series forecasting problems so difficult compared to other sequence learning tasks or IID Machine Learning problems. First some motivating baselines illustrating the difficulty of the forecasting problem. I ran a series of models statistical (Naive, AutoARIMA, Theta, MSTL, Seasonal-Naive) linear/transformer neural (DLinear, NLinear, PatchTST) a gradient-boosted tree (LightGBM) zero-shot foundation models (Chronos, TimesFM, TTM) “AI” LLM prompted to continue numeric sequence (Claude Opus, Haiku) against the benchmark datasets m4_hourly,

saved by

• vain glory
• nicolas
• Akira Yoshiyama

related reading

• Against Time-Series Foundation Modelsshakoist.substack.com
• ML Blog - Everything you need to know about Time Series Forecastingmrtunguyen.github.io
• Time Series - From Analyzing the Past to Predicting the Future | Towards Data Sciencetowardsdatascience.com
• TimeGPT-1 | Hacker Newsnews.ycombinator.com
• A History of Large Language Modelsgregorygundersen.com
• The Unreasonable Effectiveness of Recurrent Neural Networkskarpathy.github.io
• [2310.07820] Large Language Models Are Zero-Shot Time Series Forecastersarxiv.org
• An introduction to time series forecasting | InfoWorldinfoworld.com
• 2402.03885arxiv.org
• Is Capability a Liability? More Capable Language Models Make Worse Forecasts When It Matters Mostarxiv.org
• Introduction to Time Series Analysis — Time Series Analysis Handbookphdinds-aim.github.io
• Real-time machine learning: challenges and solutionshuyenchip.com