Feature exploration: Microsoft
6659 labelled trading days and 20 engineered features. Each feature uses only information available at that day's close; the target is whether the next close is higher. This page asks, before any model is trained: is there any signal here at all?
Base rate
51.1% of days were followed by a rise
The two classes are nearly balanced, so a useless model scores about 50%. Statistics on this page use the full history, including the later test period. They describe the data and play no part in model selection.
Is tomorrow predictable?
The core difficulty, in two charts
Takeaway 4 of 20 lags fall outside the band, but the largest is only |ρ| = 0.064 (lag 1), which explains 0.41% of the variation. Volatility clustering also makes the simple band too narrow, so even these are weaker than they look.
Takeaway Excess kurtosis is 9.8 (a normal distribution has 0): moves bigger than 3σ happen on 1.6% of days against 0.3% for a normal curve.
Feature signal
Does each feature carry information about the next day?
Takeaway 6 of 20 features clear the noise band; the strongest (oc_change) has ρ = -0.066, so it explains well under 1% of the variation in next-day returns.
How to read Pick a feature from the dropdown. A predictive feature would show a clear slope from Q1 to Q10; a flat line inside the error bars means no edge.
Redundancy
Which features duplicate each other
Takeaway Tight blue blocks (the SMA and EMA ratios, for example) are near-duplicates. That's why linear models need regularisation, and why permutation importance can look small for each twin.