Auto-Weighted Blend
In plain terms
Per-ticker automatically learned blend weights with out-of-sample gating. Instead of equal-weighting the top-3 strategies, an optimizer searches over how many to blend, which to pick, how much weight to give each, what regime to fire in, and when to kill the position on a drawdown; only blends that survive fold-wise out-of-sample Sharpe checks, permutation testing on a held-out final year, and a trial-count penalty on the Sharpe ratio are persisted.
How it works
The existing the best few blenders (auto-blend, production-blend, production blends, ensemble, ensemble-v2, the ensemble) all do rule-based component selection (the best few by grade/sharpe, equal-weight). Per ticker we typically have 30-150 discovered strategy_experiments rows with retained _position_series; an an automatic search study learns (K, rank-metric, Dirichlet weights, regime gate, drawdown kill-switch) directly against the blended objective. Gates: (1) fold-mean OOS Sharpe > 0 over 6 contiguous folds with a 2-day boundary shrink (a simplified stand-in for cross-validation: no combinatorial train/test splits, no purged training set, weights are not re-fit per fold); (2) 500-shuffle random-shuffle permutation on the disjoint ~252-bar holdback slice of the SYNTHETIC series, p < 0.05 (the missing check in rule-based blenders; lag-aligned and faithful); (3) trial-count Sharpe deflation SR*sqrt(1 - ln(n_trials)/N), a Bonferroni-style proxy for the Deflated Sharpe Ratio that is near-inert at the default n_trials=80 (factor ~0.99); (4) n_trades >= 20. A trial that fails any gate returns -inf; only winners persist as saved strategies rows competing in the picker pool.
Data dependencies
- Daily prices
Adjusted-close OHLCV for every US-listed ticker; primary price feed.
- Strategy experiments
A data feed this strategy reads, refreshed on its normal schedule.
- Market regime daily
A data feed this strategy reads, refreshed on its normal schedule.
Expected edge
- Tested over
- Per-ticker holdback (last ~252 bars)
Learned blend weights with statistically-validated synthetic series should dominate equal-weight the best few on Sharpe and luck-adjusted scoring — the rule-based blenders have never random-shuffle-tested their synthesized series.
Related families
Instead of trying to optimize weights across sub-signals (which overfits), just equal-weight K canonical price signals. Hard to beat out-of-sample.
Instead of equal-weighting our internal sub-signals, cluster them by how correlated their returns are and give each cluster a proportionate slice of the risk budget. Stable, classic Lopez de Prado allocator that systematically beats naive 1/N out-of-sample.
When no single strategy clearly wins, we combine the best three TA strategies with the best three alpha-family strategies — half-and-half — and use that blend as the champion. On stocks where every individual strategy was failing, the blend rescued 2 out of 3 of them.
Take the user's best historical strategy for this ticker, then stack additional regime filters (VIX, sector trend, short flow) on top to clean up its bad bars.
Explore Auto-Weighted Blend on alphactor.ai
See which tickers this family is currently firing on, with live signals and rankings.