Choosing Prediction Targets for Technical Indicator Models

Summary

This article argues that financial machine learning performance depends heavily on how the prediction target is defined. Unlike a fixed medical label, a trading target may represent returns, price appreciation, volatility, drawdown, or relative asset movement over different horizons. The author proposes testing alternative targets as a way to improve models when progress stalls, rather than assuming that more data or greater model complexity will solve the problem.

The practical example builds a market dataset from OHLC prices and moving averages, along with changes across time horizons and price levels. The article compares successive modeling approaches and reports that modeling a moving average at multiple horizons and trading its implied slope performed best among the described versions. Expanding the data reduced profitability, while a nonlinear learner on the larger dataset did better than that expanded-data version but did not match the earlier approach. These are backtest results from the author's example, so they do not establish that the target or strategy generalizes across markets or periods.

Key ideas

• Financial models can be constrained by the target definition as well as by data quality or model choice.
• Possible targets include returns, volatility, drawdown, price changes, and relative asset movement.
• The example derives features from OHLC prices, moving averages, and changes across bars and price levels.
• The reported best version modeled moving average behavior at multiple horizons and traded its implied slope.
• The article's performance comparisons come from a specific backtest and do not establish broad generality.

Tags

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