Alpha factors are designed to extract signals from data to predict asset returns for a given investment universe over the trading horizon. A factor takes on a single value for each asset when evaluated, but may combine one or several input variables. The process involves the steps outlined in the following figure:

The Research phase of the trading strategy workflow includes the design, evaluation, and combination of alpha factors. ML plays a large role in this process because the complexity of factors has increased as investors react to both the signal decay of simpler factors and the much richer data available today.

The development of predictive alpha factors requires the exploration of relationships between input data and the target returns, creative feature-engineering, and the testing and fine-tuning of data transformations to optimize the predictive power of the input.

The data transformations range from simple non-parametric rankings to complex ensemble models or deep neural networks, depending on the amount of signal in the inputs and the complexity of the relationship between the inputs and the target. Many of the simpler factors have emerged from academic research and have been increasingly widely used in the industry over the last several decades. 

To minimize the risks of false discoveries due to data mining and because finance has been subject to decades of research that has resulted in several Nobel prizes, investors prefer to rely on factors that align with theories about financial markets and investor behavior. Laying out these theories is beyond the scope of this book, but the references will highlight avenues to dive deeper into this important framing aspect of algorithmic trading strategies.

To validate the signal content of an alpha factor candidate, it is necessary to obtain a robust estimate of its predictive power in environments representative of the market regime during which the factor would be used in a strategy. Reliable estimates require avoiding numerous methodological and practical pitfalls, including the use of data that induces survivorship or look-ahead biases by not reflecting realistic PIT information, or the failure to correct for bias due to multiple tests on the same data.

Signals derived from alpha factors are often individually weak, but sufficiently powerful when combined with other factors or data sources, for example, to modulate the signal as a function of the market or economic context.