Skip to content
Stratum

Research Layer

Last updated: 2026-05-19

Overview

The research layer sits above the gold layer and answers: “which signals actually work, and how should I combine them into a strategy?” It consists of three components, deployed incrementally:

Gold signals ──┐
               ├──► Signal Validation (walk-forward)  ──► Signal Scorecard     [IMPLEMENTED]
               ├──► Meta-Model (ML combination)       ──► Combined Alpha Score [PLANNED #1]
               └──► Backtest Engine                   ──► Strategy PnL + Risk  [PLANNED #2]

1. Walk-Forward Signal Validation (Implemented)

Purpose

Continuously validate whether gold signals have predictive power over forward equity returns. Detects signal decay, regime sensitivity, and overfitting — so you know which signals to trust before trading on them.

Methodology

Expanding window approach:

|------------ train (>= 252 days) ------------|-- test (63 days) --|
                                              ^
                                              window slides forward by 21 days
  1. Start with minimum 252 trading days of history
  2. Test the next 63 days (one quarter)
  3. Slide the window forward by 21 days
  4. Repeat until end of available data

Metrics per window:

MetricDefinitionThreshold
IC (Information Coefficient)Spearman rank correlation: signal vs. forward return|IC| > 0.03 is meaningful
Hit Rate% of days where signal direction matched return direction> 55% is meaningful
t-statisticStatistical significance of IC|t| > 1.96 (95% confidence)
Regime ICIC computed within each regime subset (hiking/cutting/stress/etc.)Same thresholds

Signal health classification:

HealthCriteriaAction
ActiveLatest IC > 0.03 AND statistically significantUse in strategies
WeakeningHistorical avg IC > 0.02 but latest not significantMonitor, reduce weight
DeadNo evidence of predictive powerDo not trade on this

What Gets Validated

All gold signal columns across three horizons (5d, 20d, 60d forward SPY return):

Yield Curve signal:

  • DGS30, DGS10, DGS2, T10Y2Y levels
  • dgs10_dgs2_spread, dgs30_dgs10_spread
  • t10y2y_ma20, t10y2y_ma60
  • dgs10_momentum (20d, 60d), dgs30_momentum (20d, 60d)
  • Regime: fed_cycle (hiking/cutting/hold)

Risk Composite signal:

  • risk_score, risk_score_smoothed
  • VIXCLS_zscore, DTWEXBGS_zscore
  • spy_momentum_20d_zscore, hyg_spy_ratio_zscore
  • Regime: regime (risk_on/neutral/risk_off)

Credit Leverage signal:

  • credit_stress_score, credit_stress_smoothed
  • BAMLH0A0HYM2, BAMLC0A4CBBB, hy_bbb_spread_diff
  • hy_spread_momentum_20d, bbb_spread_momentum_20d
  • margin_debt_pct_chg_3m
  • Regime: credit_regime (stress/elevated/neutral/benign)

Infrastructure

  • ETL: etl/research/walk_forward_validation.py
  • Compute: AWS Glue (3x G.1X workers)
  • Schedule: Weekly, Sunday 08:00 UTC
  • Output: s3://{research_bucket}/validation/signal={name}/ + validation/scorecard/
  • Trigger: EventBridge emits SignalValidationComplete when scorecard is ready

How to Read the Scorecard

Query via Athena:

SELECT signal_name, horizon_days, latest_ic, latest_hit_rate, signal_health
FROM stratum_research_{env}.signal_scorecard
WHERE signal_health != 'dead'
ORDER BY latest_ic DESC;

Key questions it answers:

  • “Which signals are currently predictive at the 20-day horizon?”
  • “Has credit_stress_score lost its edge since the regime shifted to ‘hold’?”
  • “Which signals work in a hiking cycle but die in a hold?“

2. ML Signal Combination — Meta-Model (Planned)

GitHub: Issue #1

Concept

Individual signals are noisy and fire at different timings. The meta-model learns the optimal weighting and timing across all gold signals, producing a single daily probability score.

Architecture

  • Model: LightGBM (interpretable, handles mixed frequencies)
  • Features: All gold signal columns aligned on date
  • Target: Forward SPY returns at multiple horizons, or binary drawdown indicator (> 5% in next 20 days)
  • Training: Walk-forward (expanding window) — never uses future data
  • Output: Daily combined_alpha score + feature importances

Why Not a Simple Average?

  • Signals have different optimal horizons (credit stress leads by 2-4 weeks; VIX is contemporaneous)
  • Signal importance changes by regime (rates momentum matters more during hiking)
  • Non-linear interactions (credit stress + rising rates = multiplicative risk)

3. Strategy Backtesting Framework (Planned)

GitHub: Issue #2

Concept

Translate validated signals into simulated portfolio decisions with realistic constraints.

Architecture

  • Engine: Event-driven loop (iterates day by day, no lookahead)
  • Inputs: Gold signals + meta-model + Yahoo price histories
  • Assets: SPY, TLT, GLD, cash (expandable)
  • Constraints: Transaction costs, slippage, position limits
  • Output: Equity curve, drawdown, Sharpe, Calmar, turnover

Example Strategies to Test

  1. Risk-off on credit stress: Go 50% cash when credit_stress_score > 1.0
  2. Rates + leverage squeeze: Short equities when dgs30_momentum_60d > 50bps AND credit_regime == "stress"
  3. Combined alpha threshold: Size positions proportional to meta-model combined_alpha

Dependency Chain

Walk-forward validation (which signals to trust)
        │
        ▼
Meta-model (how to combine trusted signals)
        │
        ▼
Backtester (what would this combination have returned)

Each layer feeds the next. Walk-forward tells you which inputs to give the meta-model. The meta-model produces the score the backtester trades on.