Signal Research

This skill provides the systematic methodology for discovering, engineering, evaluating, and combining predictive trading signals (alpha signals) using quantitative methods. It covers the full research pipeline — from raw data to deployable signal — including factor model construction, feature engineering, signal evaluation metrics (Information Coefficient, Information Ratio, decay analysis), combinatorial signal testing with overfitting safeguards, time-series cross-validation, and crypto-specific signal discovery. This is the research layer that feeds into trading strategies.

When to Use This Skill

• When searching for new predictive signals (alpha discovery)
• When constructing factor models (single-factor or multi-factor)
• When engineering features from raw market/on-chain data
• When evaluating whether a signal has genuine predictive power
• When analyzing how quickly a signal decays over time
• When combining multiple signals into a composite model
• When applying proper cross-validation for time-series financial data
• When quantifying crypto-specific signals (funding rate, sentiment, on-chain metrics)

What This Skill Does

1. Alpha Signal Discovery: Systematic process for generating and screening candidate signals
2. Factor Model Construction: Builds single-factor and multi-factor models with proper orthogonalization
3. Feature Engineering: Transforms raw data into predictive features (returns, z-scores, ratios, ranks)
4. Signal Evaluation: Measures predictive power using IC, IR, turnover, and decay rate
5. Overfitting Prevention: Applies multiple testing corrections, walk-forward validation, and out-of-sample discipline
6. Signal Combination: Combines signals using ranking, z-scoring, and weighted composites
7. Crypto Signal Specialization: Evaluates funding rate signals, social sentiment quantification, and on-chain metric factors

How to Use

Alpha Discovery

Discover new alpha signals for BTC using price, volume, and funding rate data over the past 2 years

Factor Model

Build a multi-factor model for crypto top-20 assets using momentum, value, and volatility factors

Signal Evaluation

Evaluate the predictive power of this signal: 7-day funding rate z-score for ETH perpetual futures

Feature Engineering

Engineer trading features from raw OHLCV + on-chain data for [token]. What transformations add predictive value?

Signal Decay

How quickly does the momentum signal (20-day return) decay for BTC? What's the optimal holding period?

Data Sources

With MCP/CLI tools connected:

• yFinance MCPs (tooyipjee, maxscheijen, Adity-star) — Historical OHLCV, returns computation, fundamental data
• Empyrical MCP — Performance metrics, risk-adjusted returns, factor analysis support
• CoinGecko MCP / CoinGecko Price MCP — Crypto price history, market cap, volume, social metrics
• Binance MCP (TermiX, snjyor) — Funding rates, open interest, order book depth, liquidation data
• Crypto Sentiment MCP — Social sentiment scores, Fear & Greed index, social volume
• DeFiLlama — TVL data, fee revenue, protocol metrics for on-chain factor construction
• Whale Tracker MCP — Whale flow data for exchange flow factors

Without tool access: Ask the user to provide:

1. Historical price data (OHLCV) for the asset(s) — minimum 252 trading days (1 year), ideally 500+
2. Any additional data series they want to test as signals (funding rates, sentiment scores, on-chain metrics)
3. The universe of assets (single asset or cross-sectional)
4. Target holding period (1 day, 1 week, 1 month)
5. Forward return data for evaluating signal predictions

Proceed with analysis framework using provided data. Note statistical limitations of small samples.

Methodology

Step 1: Define the Research Question

Before generating signals, precisely define what you are trying to predict.

RESEARCH QUESTION FRAMEWORK:

1. TARGET VARIABLE: What are you predicting?
   - Forward 1-day return (short-term)
   - Forward 5-day return (swing trading)
   - Forward 21-day return (monthly)
   - Forward return sign (direction only)
   - Risk-adjusted forward return (Sharpe of next N days)

2. UNIVERSE: What assets?
   - Single asset (BTC, ETH, SPY, etc.)
   - Cross-sectional (top 50 crypto, S&P 500 members)
   - Single asset = time-series signal
   - Cross-sectional = relative ranking signal

3. SIGNAL TYPE: What kind of edge?
   - Momentum: Past returns predict future returns
   - Mean-reversion: Extreme moves revert
   - Value: Cheap assets outperform expensive ones
   - Quality: High-quality assets outperform low-quality
   - Sentiment: Crowd behavior predicts short-term moves
   - Flow: Money movement predicts price direction

4. HOLDING PERIOD: How long will you hold?
   - Must match the signal's predictive horizon
   - A 1-day signal used for 30-day holding = misuse

Step 2: Feature Engineering — Raw Data to Predictive Features

Transform raw data into features that may have predictive power.

Core Feature Transformations:

RETURNS (most common base feature):
  Simple return: r_t = (P_t - P_{t-1}) / P_{t-1}
  Log return: r_t = ln(P_t / P_{t-1})  [better for compounding analysis]
  N-day return: r_{t,N} = (P_t - P_{t-N}) / P_{t-N}
  Periods: 1d, 5d, 10d, 21d, 63d, 126d, 252d

VOLATILITY:
  Realized vol: σ_N = StdDev(r_t, N) × √252  [annualized]
  Parkinson vol: σ_P = √(1/(4N×ln2) × Σ(ln(H_t/L_t))²)  [uses high-low, more efficient]
  Vol ratio: σ_short / σ_long  [vol regime detection]
  Periods: 5d, 10d, 21d, 63d

Z-SCORES (standardization for cross-sectional comparisons):
  z = (x - μ) / σ
  Rolling z-score: z_t = (x_t - mean(x, N)) / std(x, N)
  Cross-sectional z: z_i = (x_i - mean(x, all assets)) / std(x, all assets)
  Common window: 21-day, 63-day rolling

RATIOS:
  Price / SMA(N): Trend deviation ratio
  Volume / SMA(Volume, N): Volume surge ratio
  ATR(short) / ATR(long): Volatility expansion/contraction ratio
  RSI: Already a ratio (0-100 scale)

RANKS (robust to outliers -- preferred for cross-sectional):
  Percentile rank: rank_i = rank(x_i) / N
  No outlier sensitivity (rank of 1000% return = same as rank of 50% return)
  Always normalize: 0 (worst) to 1 (best)

INTERACTION FEATURES:
  Momentum × Volume: High return + high volume = stronger signal
  Volatility × Direction: Low vol + positive return = quality trend
  Price vs MA × RSI: Near MA + oversold RSI = mean-reversion setup

LAGGED FEATURES:
  Signal value at t-1, t-2, ..., t-N
  Useful for: detecting signal persistence and autocorrelation
  Caution: More lags = more overfitting risk

Crypto-Specific Feature Engineering:

FUNDING RATE FEATURES:
  Raw funding rate (8h intervals)
  Cumulative funding (7d, 14d, 30d)
  Funding rate z-score: z = (FR_t - mean(FR, 30d)) / std(FR, 30d)
  Funding rate momentum: FR_t - FR_{t-7}
  Extreme funding dummy: 1 if |FR| > 0.05%, else 0

SENTIMENT FEATURES:
  Fear & Greed Index (0-100): Contrarian signal at extremes
  Social volume z-score: Unusual social activity
  Weighted sentiment: Volume-weighted average of sentiment scores
  Sentiment divergence: Price rising but sentiment falling (or vice versa)

ON-CHAIN FEATURES:
  Exchange flow z-score: Net flow vs 30d average (see onchain-analysis skill)
  Active address momentum: 7d MA / 30d MA of daily active addresses
  NVT ratio: Market cap / Daily on-chain transaction volume
  MVRV ratio: Market cap / Realized cap (BTC/ETH specific)
  Staking ratio change: Delta in % of supply staked over 7d/30d
  Whale accumulation score: Net whale flow z-score (see onchain-analysis)

Step 3: Single-Factor Signal Evaluation

Evaluate each candidate signal individually before combining.

Information Coefficient (IC):

IC = Spearman rank correlation between signal value and forward return

IC INTERPRETATION:
  | IC Value | Interpretation | Quality |
  |----------|---------------|---------|
  | > 0.10 | Very strong | Rare -- check for overfitting |
  | 0.05-0.10 | Strong | Good single factor |
  | 0.02-0.05 | Moderate | Useful in combination |
  | 0.00-0.02 | Weak | May add marginal value in multi-factor |
  | < 0.00 | Negative or zero | Discard (or use as reverse signal if significant) |

WHY SPEARMAN (not Pearson):
  - Robust to outliers (crypto has extreme outliers)
  - Captures monotonic relationships (not just linear)
  - Works with ranked portfolios (long top decile, short bottom decile)

CALCULATION:
  For each period t:
    IC_t = SpearmanCorr(Signal_values_t, Forward_returns_{t+1 to t+H})

  Average IC = mean(IC_t across all periods)
  IC standard error = std(IC_t) / √(N_periods)
  IC t-stat = Average IC / IC standard error

  STATISTICAL SIGNIFICANCE:
  t-stat > 2.0 (p < 0.05): Signal is statistically significant
  t-stat > 2.5: More robust significance
  t-stat > 3.0: Strong evidence of predictive power

Information Ratio (IR):

IR = Mean(IC) / StdDev(IC)

IR = how consistently the signal predicts (Sharpe ratio of IC)

  | IR Value | Interpretation |
  |----------|---------------|
  | > 0.5 | Excellent -- very consistent signal |
  | 0.3-0.5 | Good -- reliable signal |
  | 0.1-0.3 | Moderate -- usable but inconsistent |
  | < 0.1 | Poor -- too noisy to rely on |

A signal can have a decent IC but poor IR if it only works in certain market regimes.

Turnover:

Turnover = fraction of portfolio that changes when signal is re-evaluated

Turnover = Σ|weight_{t} - weight_{t-1}| / 2

LOW TURNOVER (< 20% monthly): Cheap to trade, long-term factors
HIGH TURNOVER (> 100% monthly): Expensive, must overcome transaction costs

NET IC = IC - (Turnover × Transaction_Cost_per_unit)

If Net IC < 0: Signal is profitable before costs but unprofitable after costs → DISCARD

TRANSACTION COST ESTIMATES:
  | Market | One-way Cost | Round-trip |
  |--------|-------------|-----------|
  | US equities (liquid) | 5 bps | 10 bps |
  | US equities (small) | 15 bps | 30 bps |
  | Crypto (BTC/ETH, CEX) | 5-10 bps | 10-20 bps |
  | Crypto (altcoin, CEX) | 10-30 bps | 20-60 bps |
  | Crypto (DEX, major) | 30-50 bps | 60-100 bps |
  | Crypto (DEX, small) | 50-300 bps | 100-600 bps |

Step 4: Signal Decay Analysis

Measure how quickly a signal loses predictive power over time.

SIGNAL DECAY METHODOLOGY:

1. Compute IC at different forward horizons:
   IC_1d = IC using 1-day forward returns
   IC_2d = IC using 2-day forward returns
   IC_5d = IC using 5-day forward returns
   IC_10d = IC using 10-day forward returns
   IC_21d = IC using 21-day forward returns

2. Plot IC vs horizon → decay curve

3. Fit exponential decay: IC(h) = IC_0 × e^(-λh)
   - IC_0: signal strength at generation
   - λ: decay rate (higher = faster decay)
   - Half-life = ln(2) / λ

4. OPTIMAL HOLDING PERIOD = horizon where IC is maximized
   or = half-life of the decay curve

TYPICAL DECAY RATES:

| Signal Type | Half-Life | Optimal Holding |
|------------|-----------|-----------------|
| Momentum (short-term, 5d return) | 3-7 days | 5-10 days |
| Momentum (medium, 21d return) | 10-30 days | 15-30 days |
| Momentum (long, 126d return) | 30-90 days | 60-120 days |
| Mean-reversion (RSI extreme) | 1-3 days | 1-5 days |
| Funding rate extreme | 1-5 days | 2-7 days |
| Sentiment extreme | 1-3 days | 1-5 days |
| On-chain flow signal | 3-14 days | 5-21 days |
| Value (NVT, P/E) | 30-90 days | 30-90 days |

IF A SIGNAL HAS NO DECAY (flat IC across horizons):
  - May be a structural alpha (rare and valuable)
  - Or may be a data artifact -- investigate carefully
  - Test on out-of-sample data

IF A SIGNAL DECAYS IN < 1 DAY:
  - Only usable for high-frequency trading
  - Transaction costs likely eat the edge
  - Requires low-latency execution infrastructure

Step 5: Cross-Validation for Time Series

Standard k-fold cross-validation is INVALID for time series because it leaks future information. Use these methods instead.

METHOD 1: WALK-FORWARD VALIDATION (Recommended)

  Train: [────────────] (e.g., days 1-252)
  Test:       [────]    (e.g., days 253-273)

  Then slide forward:
  Train: [──────────────] (e.g., days 22-273)
  Test:         [────]    (e.g., days 274-294)

  Parameters:
  - Training window: 252 trading days (1 year) minimum
  - Test window: 21 trading days (1 month) typical
  - Step size: 21 days (non-overlapping test periods)
  - Total walks: (Total days - Training window) / Step size

METHOD 2: EXPANDING WINDOW

  Train: [──]          (days 1-252)
  Test:     [──]       (days 253-273)

  Train: [────]        (days 1-273)
  Test:       [──]     (days 274-294)

  Train: [──────]      (days 1-294)
  Test:         [──]   (days 295-315)

  Advantage: Uses all available training data
  Disadvantage: Earlier data may be less relevant (regime changes)

METHOD 3: PURGED K-FOLD (for cross-sectional signals)

  Standard k-fold BUT:
  1. Sort by time
  2. Create a GAP (purge window) between train and test folds
  3. Gap size = maximum holding period + 1
  4. Prevents information leakage from overlapping return horizons

  Example (5-fold, 5-day forward returns):
  Fold 1: Train [all except fold 1], Gap [5 days each side], Test [fold 1]
  The 5-day gap ensures no return period in train overlaps with test

VALIDATION RULES:
  - NEVER use standard k-fold for time series data
  - NEVER include future data in training (even indirectly via normalization)
  - Test period should cover at least 2 full market cycles (bull + bear)
  - Minimum 100 test-period observations for statistical validity
  - Report ALL walk-forward results (not just the best)

Step 6: Multi-Factor Model Construction

Combine individually validated signals into a composite model.

Factor Combination Methods:

METHOD 1: EQUAL-WEIGHT Z-SCORE COMBINATION (Simplest)

  1. For each signal, compute cross-sectional z-score at each time t
  2. Composite = mean(z-score_1, z-score_2, ..., z-score_K)
  3. Rank assets by composite score
  4. Long top quintile, short bottom quintile (or just long top)

  Pros: Robust, no estimation error, hard to overfit
  Cons: Ignores signal strength differences

METHOD 2: IC-WEIGHTED COMBINATION

  1. Compute rolling IC for each signal (trailing 63 days)
  2. Weight_k = max(IC_k, 0) / Σ max(IC_j, 0)
  3. Composite = Σ(Weight_k × z-score_k)

  Pros: Adapts to which signals are currently working
  Cons: Slightly more prone to overfitting than equal-weight

METHOD 3: REGRESSION-BASED (Most Powerful, Most Dangerous)

  Forward_return = β_1 × Signal_1 + β_2 × Signal_2 + ... + β_K × Signal_K + ε

  Estimated via: OLS, Ridge, Lasso, or Elastic Net
  - Ridge (L2): Shrinks coefficients, handles multicollinearity
  - Lasso (L1): Sets weak signals to exactly zero (feature selection)
  - Elastic Net: Combination of Ridge + Lasso

  CRITICAL: Estimate coefficients ONLY on training data
  - Use walk-forward to estimate β on training, test on holdout
  - Re-estimate coefficients every rebalance period
  - Regularization is MANDATORY (unregularized OLS will overfit)

METHOD 4: RANK-BASED COMBINATION (Most Robust)

  1. For each signal, rank all assets (0 = worst, 1 = best)
  2. Composite rank = mean(rank_1, rank_2, ..., rank_K)
  3. Advantages: No parametric assumptions, robust to outliers

Factor Orthogonalization:

PROBLEM: Signals often overlap (e.g., momentum and trend-following are correlated)
SOLUTION: Orthogonalize factors to remove redundancy

PROCESS:
  1. Compute correlation matrix of all candidate signals
  2. If two signals have |correlation| > 0.7: they are redundant
  3. Keep the one with higher IC, discard or residualize the other
  4. Residualization: Signal_B_orth = Signal_B - β × Signal_A
     (regress B on A, use residuals as the orthogonalized B)
  5. Maximum factors in a model: 5-7 (more = overfitting risk)
  6. Each factor should capture a DIFFERENT source of return

FACTOR CATEGORY DIVERSIFICATION:
  Include at most 1-2 signals from each category:
  - Momentum (1 signal: e.g., 21d return)
  - Value (1 signal: e.g., NVT z-score)
  - Volatility (1 signal: e.g., vol ratio)
  - Flow (1 signal: e.g., exchange flow z-score)
  - Sentiment (1 signal: e.g., Fear & Greed z-score)

Step 7: Overfitting Prevention and Multiple Testing Correction

The biggest risk in signal research is finding patterns that do not generalize.

OVERFITTING WARNING SIGNS:

1. Signal works perfectly in-sample but fails out-of-sample
2. Signal requires very specific parameter values (e.g., RSI(13) works but RSI(12) and RSI(14) don't)
3. Signal has extremely high IC (>0.15) -- probably fitting noise
4. Signal needs many conditions/filters to work
5. Signal was found after testing 100+ variants

MULTIPLE TESTING CORRECTION:

When you test N signals, some will appear significant by chance.

Bonferroni correction (conservative):
  Adjusted p-value threshold = 0.05 / N
  If you tested 20 signals: need p < 0.0025 (t-stat > 3.0)
  If you tested 100 signals: need p < 0.0005 (t-stat > 3.5)

Benjamini-Hochberg (less conservative):
  1. Sort p-values from smallest to largest: p_(1) ≤ p_(2) ≤ ... ≤ p_(N)
  2. Find largest k where p_(k) ≤ (k/N) × 0.05
  3. All signals with p ≤ p_(k) are significant

HAIRCUT RULE OF THUMB:
  Reported backtest Sharpe should be haircut by:
  Adjusted Sharpe = Reported Sharpe × (1 - 0.1 × √(Number of signals tested / 10))

  Example: Tested 50 signals, best has Sharpe 2.0
  Adjusted = 2.0 × (1 - 0.1 × √5) = 2.0 × 0.776 = 1.55

STRUCTURAL SAFEGUARDS:
  1. Define your signal universe BEFORE testing (don't add signals after seeing results)
  2. Reserve 30% of data as a FINAL holdout -- never touch until the very end
  3. Use walk-forward validation (not in-sample optimization)
  4. Prefer simple models (fewer parameters = less overfitting)
  5. Require economic rationale for every signal (not just statistical significance)
  6. Test on multiple assets / time periods -- signal should generalize

Step 8: Signal Deployment Decision Framework

After evaluation, decide whether a signal is ready for live trading.

SIGNAL READINESS CHECKLIST:

| Criterion | Threshold | Pass/Fail |
|-----------|-----------|-----------|
| Average IC | > 0.02 (single asset) or > 0.03 (cross-sectional) | |
| IC t-statistic | > 2.5 (after multiple testing correction) | |
| Information Ratio | > 0.2 | |
| Out-of-sample IC | > 50% of in-sample IC | |
| Walk-forward Sharpe | > 0.5 (after transaction costs) | |
| Signal decay half-life | > 2× rebalance frequency | |
| Turnover (monthly) | Low enough that costs < 50% of gross alpha | |
| Economic rationale | Clear, documented | |
| Works in multiple periods | Bull + bear + sideways | |
| Survives multiple testing correction | Yes | |

DECISION:
  All criteria pass → Deploy at small size, monitor for 3 months
  1-2 criteria fail → Refine signal, retest
  3+ criteria fail → Discard signal or return to feature engineering

DEPLOYMENT SIZING:
  New signal: 25% of full intended allocation
  After 3 months positive: 50% of full allocation
  After 6 months positive: 75% of full allocation
  After 12 months positive: full allocation
  At any point if signal degrades: reduce to 25% or halt

Anti-Patterns

DO NOT do these — they are the most common signal research failures:

1. Data snooping / p-hacking: Testing hundreds of signal variations and reporting only the best one. Apply multiple testing corrections and require out-of-sample validation.

2. Lookahead bias: Using information that would not have been available at the time of the signal. Example: Using a company’s annual earnings report data on the day it was filed, not the day it was released. Always use point-in-time data.

3. Survivorship bias: Only analyzing assets that exist today. Crypto tokens that went to zero, stocks that were delisted — ignoring them inflates backtest performance. Use full historical universe.

4. Overfitting to parameters: If your signal only works with RSI(13) and not RSI(12) or RSI(14), it is overfit. Good signals are robust to small parameter changes.

5. Ignoring transaction costs: A signal with IC of 0.05 but 200% monthly turnover is worthless in crypto after DEX fees. Always compute net-of-cost IC and Sharpe.

6. Confusing correlation with causation: Just because funding rate predicts returns does not mean funding rate causes returns. Understand the mechanism — funding rate extremes attract arbitrageurs who push price.

7. Using standard k-fold on time series: This leaks future information into training data. Always use walk-forward, expanding window, or purged k-fold.

8. Too many factors in one model: More than 5-7 factors in a model virtually guarantees overfitting. Each factor must contribute independent, orthogonal predictive power.

Validation Checkpoints

Input Validation

• Research question clearly defined (target variable, universe, holding period)
• Sufficient historical data (minimum 252 days; 500+ preferred; crypto: 24/7 data)
• Data is point-in-time (no lookahead bias)
• Universe includes delisted/dead assets (no survivorship bias)
• Data is clean (no gaps, outliers handled, adjusted for splits/forks)
• Training/test split defined BEFORE any signal testing begins
• Multiple testing correction method selected

Signal Evaluation Validation

• IC computed using Spearman rank correlation
• IC t-statistic reported with statistical significance
• Information Ratio computed and above minimum threshold
• Signal decay curve plotted with half-life calculated
• Turnover computed and net-of-cost IC is positive
• Walk-forward validation used (not in-sample optimization)
• Out-of-sample IC is at least 50% of in-sample IC
• Multiple testing correction applied

Model Validation

• Factors are not highly correlated (|r| < 0.7 pairwise)
• Economic rationale documented for every factor
• Model has at most 5-7 factors
• Walk-forward Sharpe is positive after transaction costs
• Model tested across multiple market regimes
• Final holdout data confirms results

Example

User: “Evaluate whether the 7-day funding rate z-score is a predictive signal for BTC forward 5-day returns.”

Analysis:

## Signal Evaluation: BTC Funding Rate Z-Score → 5-Day Forward Return

### Signal Definition
- Signal: z_t = (FR_7d_avg_t - mean(FR_7d, 90d)) / std(FR_7d, 90d)
  - FR_7d_avg = average 8-hour funding rate over past 7 days
  - Z-scored against trailing 90-day distribution
- Hypothesis: Extreme positive funding (overleveraged longs) predicts
  negative 5-day returns. Extreme negative funding (overleveraged shorts)
  predicts positive 5-day returns. Signal is CONTRARIAN.
- Economic rationale: Extreme funding rates create carry costs that force
  position unwinds. Arbitrageurs also exploit extreme funding by taking
  the opposing side, pushing price in the contrarian direction.

### Data
- Asset: BTC/USDT perpetual (Binance)
- Period: Jan 2021 -- Dec 2024 (4 years, ~1,460 daily observations)
- Training: Jan 2021 -- Dec 2023 (3 years)
- Test: Jan 2024 -- Dec 2024 (1 year, fully out-of-sample)
- Funding rate: 8-hour intervals, averaged to daily

### Feature Engineering
- Raw signal: 7-day average funding rate
- Transformation: 90-day rolling z-score
- Signal value: Negated (multiply by -1) because signal is contrarian
  - High z → overlevered longs → bearish → we want negative signal
  - Low z → overlevered shorts → bullish → we want positive signal

### Results: Information Coefficient

| Period | Mean IC | IC Std Dev | IC t-stat | IR |
|--------|---------|-----------|-----------|-----|
| Full sample (2021-2024) | 0.047 | 0.12 | 3.24 | 0.39 |
| Training (2021-2023) | 0.052 | 0.13 | 2.98 | 0.40 |
| Test (2024) | 0.038 | 0.10 | 2.71 | 0.38 |
| Bull regime (2021, 2024 Q4) | 0.061 | 0.14 | 2.42 | 0.44 |
| Bear regime (2022) | 0.039 | 0.11 | 2.18 | 0.35 |
| Sideways regime (2023) | 0.041 | 0.12 | 2.05 | 0.34 |

### Signal Decay Analysis

| Forward Horizon | IC | Significance |
|----------------|-----|-------------|
| 1-day | 0.021 | Weak |
| 2-day | 0.032 | Moderate |
| 3-day | 0.041 | Moderate |
| 5-day | 0.047 | Strong (peak) |
| 7-day | 0.039 | Moderate |
| 10-day | 0.028 | Moderate |
| 14-day | 0.015 | Weak |
| 21-day | 0.005 | Negligible |

Half-life: ~8 days
Optimal holding period: 5-7 days
Signal is largely exhausted by day 14

### Turnover Analysis
- Signal rebalance frequency: Weekly (every 5 days)
- Average position change per rebalance: 35% (one-way turnover)
- Transaction cost assumption: 10 bps round-trip (Binance taker)
- Monthly turnover cost: ~30 bps
- Gross monthly alpha (from IC): ~1.4%
- Net monthly alpha: ~1.1% → Viable after costs

### Walk-Forward Validation
- 3-year training, 1-year test
- Walk-forward Sharpe (net of costs): 1.42
- Out-of-sample IC retention: 73% of in-sample (0.038 / 0.052)
- Signal works in all 3 regimes tested

### Multiple Testing Correction
- Signals tested in this research batch: 8
- Bonferroni threshold: p < 0.05/8 = 0.00625 → t-stat > 2.73
- Full-sample t-stat: 3.24 > 2.73 → PASSES Bonferroni correction
- Test-set t-stat: 2.71 → borderline (close to 2.73)
  → Signal is significant but not overwhelmingly so on test alone

### Signal Readiness Checklist

| Criterion | Value | Pass? |
|-----------|-------|-------|
| Average IC > 0.02 | 0.047 | Yes |
| IC t-stat > 2.5 (corrected) | 3.24 (full), 2.71 (test) | Yes (marginal on test) |
| IR > 0.2 | 0.39 | Yes |
| OOS IC > 50% of IS | 73% | Yes |
| Walk-forward Sharpe > 0.5 | 1.42 | Yes |
| Half-life > 2× rebalance | 8d > 2×5d? No (8 < 10) | Marginal |
| Net-of-cost alpha positive | 1.1% monthly | Yes |
| Economic rationale | Clear (leverage unwind) | Yes |
| Works across regimes | Bull, bear, sideways | Yes |
| Survives multiple testing | Bonferroni p < 0.00625 | Yes |

### Recommendation
APPROVED for deployment at 25% allocation. Signal has genuine predictive
power for BTC 5-day returns with economic rationale (leverage unwind effect).
The contrarian funding rate signal is strongest at 5-7 day horizons with a
half-life of ~8 days.

Caution: Marginal significance on the test set alone suggests this signal
should be combined with other factors (momentum, exchange flow) rather than
traded in isolation. See [Risk Management](/trading-skills/risk-management/) for position sizing.

Deploy: 25% size for 3 months → evaluate → scale if Sharpe holds > 1.0.