Using Multiple Forecast Horizons to Interpret Machine Learning Signals

Summary

The article argues that a model’s forecasts may be difficult to compare directly with observed prices because predictions and targets can occupy differently transformed coordinate spaces. It explains this idea through a geometric interpretation of linear regression and singular value decomposition, describing model fitting as rotations, scaling, and projection of target information onto input features. The discussion presents this as a source of irreducible error, especially when market inputs do not fully explain the target.

Its proposed trading approach compares a model’s own forecasts at different future horizons and uses their relative slope to determine direction, rather than comparing a forecast with current price. In a reported three-year backtest, this interpretation improved net profit, Sharpe ratio, and the share of profitable trades over the article’s one-step control, without changing the model. The evidence is specific to the described setup and historical period; it does not establish that the method generalizes across markets, models, or execution conditions.

Key ideas

• The article frames linear regression as a geometric transformation that projects target information into the space spanned by inputs.
• When inputs do not fully explain a market target, forecasts may retain a mismatch with observed values.
• The proposed signal compares forecasts across future horizons and trades according to their relative slope.
• A reported historical test found better metrics than the one-step control under the same model.
• The results are limited to the article’s particular data, strategy, and backtest period.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.