📊 When several models forecast a time series, traditional metrics such as MSE or RMSE sometimes fail to clearly separate which one is better. The author proposes viewing the problem through information theory.
Instead of measuring only the distance between predictions and actual values, the approach analyzes the entropy of residuals. If residuals resemble white noise, the model has captured much of the predictable information. This measure can distinguish models that look equal under conventional error metrics.
The article applies the idea to inflation forecasts and combines models in an entropy-weighted ensemble. It also uses Granger causal networks to understand relationships among CPI, producer prices, savings, and inventories.
⚖️ The result is not a definitive solution: the experiment is preliminary, depends on user-selected thresholds, and needs more models and optimization. However, it shows that changing the metric or geometry of a problem can reveal hidden differences.
💡 Explanation in a nutshell#
An ensemble combines several opinions to produce a more robust forecast. Entropy measures how much information remains unexplained in the errors. If a new measurement finds signals that MSE misses, it can help assign better weights to each model.
More information at the link 👇
