The Acadian reading on AI/ML for stock selection, which is based on Piotroski’s model of 9 binary variables to forecast stock return outcomes, contains the following passages:
This is an appealing setting in which to apply ML for several reasons. First, we have intuition that the selected financial characteristics may help return forecasting capabilities; this isnâ€™t an uninformed fishing expedition in a low signal-to-noise-ratio environment. Second, although reducing the individual financial attributes to binary variables and then summing them is intuitive?and transparent, we have no reason to think that it fully exploits their predictive value. Third, there is no theory or intuition to suggest how we should combine the individual quality metrics, e.g., whether to accord each equal weight or whether the predictors interact …
Implementing an ML algorithm typically involves subdividing?the historical dataset into three segments: 1) a â€œtrainingâ€ sample to estimate algorithm parameters, 2) a â€œvalidationâ€ sample to tune the algorithm and/ or select among several variants, and 3) â€œout-of- sampleâ€ historical data to evaluate the approach.
Briefly comment on how these passages reflect some of the key potential benefits, and risks, of using AI/ML techniques in investment research.