Every model in our "zoo" gets the same features, the same label (stock minus S&P 500 over the next 20 trading days) and the same walk-forward schedule: refit every 63 trading days on the trailing 756, with a 20-day gap so no label leaks into training.
The leaderboard (development period, mean daily rank IC)
Model IC t-stat
Trend rule (price vs 100-day average + 12-1 momentum) +0.0149 0.64
GRU +0.0101 0.50
Ridge regression +0.0091 0.50
LightGBM +0.0086 0.53
LSTM β0.0018 β0.09
MLP β0.0090 β0.49
Low-volatility rule β0.0251 β0.98
Twelve families in total. The best Deflated Sharpe probability, which accounts for how many we tried, was 0.16.
Three observations
Nothing is significant. A t-stat of 0.64 is well inside the noise. "Top of the table" here means "least bad".
The year-by-year numbers swing. LSTM's IC was β0.07 in 2022 and +0.14 in 2025. A model that looks brilliant in one year and broken in the next has not found structure; it has found that year.
Complexity added nothing.β¦