Per-layer weight calibration sweep

Question: For each (layer, horizon) cell in judgment.DEFAULT_HORIZON_WEIGHTS_BACKTEST, what weight maximises portfolio Sharpe? Are the shipped defaults optimal?

Methodology: Three stages:

1. Marginal sweep — scripts/sweep_layer_weights.py runs 8 layers × 3 horizons × 7 weights = 168 backtests. For each cell, all other weights held at default. Reveals each cell's marginal Sharpe response curve. Misses interactions by construction.
2. Joint refinement — stack the top marginal winners into candidate weight matrices, test, pick the best.
3. Anti-overfit validation — winner re-tested on nasdaq100 and sp600 (in-sample + OOS 2020+).

Setup: weekly rebalance, n_top=10, full production settings. sp500 2014-2026 = primary universe.

Stage 1 — marginal sweep (top winners by ΔSharpe vs baseline)

Layer	Horizon	Default w	Best w	ΔSharpe
sentiment	medium	0.15	0.35	+0.136
sentiment	long	0.15	0.60	+0.133
sentiment	short	0.15	0.60	+0.116
sector	long	0.65	0.20	+0.100
sector	medium	0.55	0.20	+0.092
sector	short	0.45	0.20	+0.086
technical	long	0.10	0.60	+0.079
turnaround	short	0.00	0.60	+0.070
technical	short	0.10	0.35	+0.064
...	...	...	...	...

Sentiment dominates; sector is meaningfully over-weighted at every horizon; news shows zero impact (as expected — news layer is zero in backtest config because yfinance can't backdate it).

Stage 2 — joint candidates DO NOT compose

Naive stacking of marginal argmaxes is catastrophic:

Candidate	ΔSharpe
A: sentiment 0.35 all-horizons	+0.098
B: sector 0.20 all	-0.019
C: A + B (sentiment + sector cut)	-0.338
D: C + technical 0.30	-0.570
E: D + turnaround/insider bumps	-0.460
F: full argmax stack	-0.746
G: conservative tilt	-0.236

Only candidate A (sentiment bump alone) beats baseline. Every other change either is null or actively destroys performance. Cutting sector — which the marginal sweep said should lift Sharpe by +0.10 — actually loses -0.34 when combined with the sentiment bump. The sector and sentiment layers carry overlapping information once both are at higher weights.

Stage 3 — refinement found a better magnitude

Sweeping the sentiment bump magnitude:

Sentiment weight (all-horizons)	Sharpe	ΔSharpe
0.15 (default)	1.461	—
0.20	1.571	+0.109
0.25	1.582	+0.120
0.30	1.570	+0.109
0.35	1.560	+0.098
0.40	1.462	+0.000
0.50	1.522	+0.060
0.70	1.299	-0.162

Optimum is 0.25, not 0.35. Response is non-monotonic — 0.40 breaks the lift, 0.70 destroys it. Per-horizon variation (0.30/0.35/0.45 etc) does no better than the uniform 0.25 setting.

Anti-overfit validation (sentiment 0.25 vs baseline)

Cell	ΔSharpe	ΔCAGR	ΔMDD
sp500 12yr (in-sample)	+0.120	+3.70 %	+4.65 %
sp500 OOS 2020+	+0.122	+4.60 %	+4.65 %
nasdaq100 12yr	+0.008	-0.83 %	-5.23 %
nasdaq100 OOS 2020+	+0.084	+2.99 %	-5.23 %
sp600 12yr	+0.009	-0.08 %	-1.70 %
sp600 OOS 2020+	-0.036	-2.61 %	-1.70 %

5/6 cells positive. In-sample +0.120 ≈ OOS +0.122 on sp500 — exceptional consistency. Nasdaq100/sp600 marginal-to-flat; no catastrophic regression on any cell.

Decision

Bump DEFAULT_HORIZON_WEIGHTS_BACKTEST sentiment row from 0.15/0.15/0.15 → 0.25/0.25/0.25. Live DEFAULT_HORIZON_WEIGHTS unchanged — live includes fundamental/news (which work live but not in backtest) so the calibration doesn't transfer cleanly. Live recalibration is a separate workstream that needs a different evaluation framework.

Two important methodology notes

1. Marginal sweeps lie about joint optima. Stacking the 24-cell argmaxes produced the WORST Sharpe (-0.746) of any candidate tested. Per-cell optima don't compose because layers carry overlapping information that re-weights perturb non-linearly. Coordinate descent (or full Bayesian opt) is the right tool; marginal scan is a first pass to identify candidates.
2. Bootstrap CIs are wide. +0.12 Sharpe is comfortably outside the cross-section noise of single backtests but still within the ~±0.30 paired-bootstrap CI for sample-path variability per the harness docstring. Treat as "probable improvement, not certain". scripts/bootstrap_recent_abs.py could confirm.

Reports written: - research/layer_weight_sweep_sp500_12yr.json (168-cell raw output)

Meta-finding: The shipped backtest defaults had sentiment too low and sector too high in isolation — but cutting sector breaks the ensemble. The robust win is a single-layer tweak (sentiment ↑) with everything else left alone. Future weight recalibration should use coordinate descent or Bayesian optimization rather than naive argmax stacking.