Implied directional accuracy for trading strategies with fixed holding periods

Machine learning forecasting studies routinely report the directional accuracy (DA) of return forecasts but rarely assess the strategy risk or the probability of failing to reach the target Sharpe ratio if the strategy based on the forecasts is deployed. For strategies with fixed holding periods, the strategy risk is equivalent to the probability that the model’s DA falls below the DA threshold implied by a given target Sharpe ratio. However, for this DA threshold, we still lack an appropriate candidate. Although investors can theoretically employ the classical threshold, which is known as implied precision, this equation assumes a constant payoff and no transaction costs, thereby understating the threshold under the fixed holding setup. Considering the dispersion of return magnitudes and transaction costs, we derive the implied DA and propose it as an appropriate candidate. The proposed threshold is neutral across models and is model-agnostic, and the drift in its parameters moves its value within the sampling error of the DA. On average, the analytical mapping behind the proposed threshold is closer to the backtests of real forecasts than that behind the implied precision, and the strategy risk computed with the implied DA is higher than that computed with the implied precision at every target Sharpe ratio. For machine learning-based strategies, an investor can then calculate the strategy risk using the proposed threshold or derive the DA that a model requires given a tolerable strategy risk. • Implied DA maps the target Sharpe ratio to the DA a fixed holding strategy requires. • Implied DA is model-agnostic and stable under out-of-sample parameter drift. • Sharpe ratios derived from implied DA match backtested Sharpe ratios on average. • Fixed holding strategies can use implied DA as the threshold for risk calculation.