Quantitative Research
Building features on a panel, scoring predictive signals, checking they add anything, and attacking your own result.
Evaluating predictive signals
Three questions turn "does this predict that?" into one number with an error bar. What picture would I see, how do I get many independent looks at it, and how do I summarise them. The result is the daily cross-sectional rank IC.
Building features on a panel
Every quantity is built per group, forward targets look forward with shift(-1), and anything whose scale differs across groups is normalised by the group's own trailing history. The eight lines that cover most cases.
Does a signal add anything beyond a known baseline?
For most interesting signals there is a boring mechanism that produces the same pattern. Three ways to check the interesting one adds anything, scored on the same protocol, with base-stack code.
Attacking your own result
A score is not a result until it has survived two attacks and a sensitivity check. Is it stable across the hold-out? Could anyone use it after turnover and costs? Does it survive removing the rows you were worried about?
Choosing a model, in order of reach
Correlation first, then OLS, then logistic for a binary target, then ridge if there are several features. What each adds, what each costs in time and explainability, and what to leave out of a short study.