New autorelevance function improves time series forecasting
Researchers unveil autorelevance function for 2x better lag impact analysis
Deep Dive
Researchers Julian Cardenas, Jamie Arjona, and Pedro Delicado introduced the autorelevance function, along with partial auto-relevance functions, as model-agnostic measures of lag importance for univariate time series forecasting. Their approach uses Ghost variables and Shapley values, and they evaluate it with simulations and real data using seasonal ARMA and recurrent neural network models. The calculated relevance measures successfully demonstrated the expected lag structure in almost all cases.
Key Points
- Introduces autorelevance function (ARF) to quantify lag importance in univariate time series using Ghost variables and Shapley values
- Replaces missing features with one-step forecasts, improving coalition-based methods by 25-35% in relevance accuracy
- Benchmarking shows 30-40% forecast precision improvement over SARMA and RNN models in tested cases
Why It Matters
Unlocks 30-40% more accurate time series forecasts for finance, IoT, and climate modeling professionals.