Backtest Expert
Design, execute, and evaluate quantitative trading strategies using historical
price data from the Finskills API. Transforms a natural-language strategy hypothesis
into a rigorous backtest with performance metrics, drawdown analysis, Fama-French
factor attribution, and walk-forward validation.
Setup
API Key required — Register at https://finskills.net to get your free key.
Header: X-API-Key: <your_api_key>
Get your API key: Register at https://finskills.net — free tier available, Pro plan unlocks real-time quotes, history, and financials.
When to Activate This Skill
Activate when the user:
- Describes a trading rule they want to test historically
- Asks "how would X strategy have performed?"
- Wants to validate a momentum, mean-reversion, or factor-based strategy
- Asks about Sharpe ratio, maximum drawdown, or strategy performance metrics
- Wants to see walk-forward or out-of-sample testing
- Asks to compare two or more strategies head-to-head
Strategy Hypothesis Formulation
Before fetching data, clarify the strategy with the user:
- Universe: Which stocks? (Single stock, S&P 500, sector, custom list)
- Signal: What is the entry trigger? (Moving average crossover, RSI level, fundamental metric, etc.)
- Entry: Long, Short, or Long/Short?
- Exit: Trailing stop, fixed target, time-based, signal reversal?
- Holding Period: Daily, weekly, monthly?
- Position Sizing: Equal weight, volatility-adjusted, Kelly?
- Benchmark: SPY (default) or sector ETF?
State the strategy in a formal grammar:
ENTRY: Buy {SYMBOL} when {condition}
EXIT: Sell when {condition} OR stop at {-X%}
HOLD: Max {N} days / bars
SIZING: {$N fixed / X% of portfolio / equal weight}
BENCH: {SPY / QQQ / custom}
PERIOD: From {YYYY-MM-DD} to {YYYY-MM-DD}
Data Retrieval — Finskills API Calls
1. Historical OHLCV for Strategy Symbols
GET https://finskills.net/v1/stocks/history/{SYMBOL}?period=5y&interval=1d
Parameters:
period: 1y, 2y, 5y, 10y, maxinterval: 1d (daily), 1wk (weekly), 1mo (monthly)
Extract: date, open, high, low, close, volume, adjustedClose
Repeat for each symbol in the strategy universe AND the benchmark (SPY or QQQ).
2. Fama-French 3-Factor Data
GET https://finskills.net/v1/free/market/fama-french
Extract: Mkt-RF (market excess return), SMB (small-minus-big), HML (high-minus-low), RF (risk-free rate)
Use for: Attribution analysis — what % of strategy return is explained by market, size, value factors
Analysis Workflow
Step 1 — Data Preparation
- Align all price series to the same trading calendar (drop non-overlapping dates).
- Use
adjustedClose for all calculations (accounts for splits and dividends).
- Compute daily returns:
r_t = (adjustedClose_t / adjustedClose_{t-1}) - 1
- Handle missing data: forward-fill up to 2 days; drop if missing > 2 consecutive days.
Step 2 — Signal Generation
Based on the strategy rules, generate entry/exit signals for each day in the backtest.
Example Signal Implementations:
Simple Moving Average Crossover:
SMA_fast = rolling_mean(adjustedClose, window=fast_n)
SMA_slow = rolling_mean(adjustedClose, window=slow_n)
Signal = 1 (Long) when SMA_fast crosses above SMA_slow
Signal = 0 (Flat) when SMA_fast crosses below SMA_slow
RSI Reversion:
RSI = 100 - 100 / (1 + avg_gain_14 / avg_loss_14)
Signal = 1 (Long) when RSI < 30 (oversold)
Signal = 0 (Flat) when RSI > 70 (overbought)
Momentum:
Momentum_N = current_price / price_N_days_ago - 1
Signal = 1 when Momentum_N > threshold
Signal = 0 otherwise
Step 3 — Portfolio Returns Calculation
strategy_return_t = Signal_{t-1} × return_t (signal from prior day, no lookahead)
benchmark_return_t = SPY daily return
Compute cumulative returns:
cumulative_return = PRODUCT(1 + strategy_return_t) - 1
Apply transaction costs (default: 0.10% per round-trip trade, adjustable):
cost_t = 0.001 × |Signal_t - Signal_{t-1}| (cost only on signal changes)
net_return_t = strategy_return_t - cost_t
Step 4 — Performance Metrics
Compute the full performance scorecard:
| Metric | Formula |
|---|
| Total Return | PRODUCT(1 + r_t) - 1 |
| Annualized Return (CAGR) | (1 + total_return)^(252/n_days) - 1 |
| Annualized Volatility | STD(r_t) × √252 |
| Sharpe Ratio | (CAGR - rf_rate) / annualized_vol |
| Sortino Ratio | (CAGR - rf_rate) / downside_vol |
| Max Drawdown | max(1 - V_t / max(V_s for s ≤ t)) |
| Calmar Ratio | CAGR / Max Drawdown |
| Win Rate | % of trading days with positive return |
| Avg Win / Avg Loss | mean positive return / mean negative return |
| Profit Factor | sum wins / sum losses |
| Alpha (vs benchmark) | CAGR_strategy - CAGR_benchmark |
| Beta | COV(r_strategy, r_benchmark) / VAR(r_benchmark) |
| Information Ratio | Alpha / tracking error |
| Number of Trades | Count of signal transitions |
| Avg Holding Period | Avg days per trade |
Step 5 — Fama-French Factor Attribution
Regress strategy excess returns against F-F factors:
r_strategy - RF = α + β₁(Mkt-RF) + β₂(SMB) + β₃(HML) + ε
Interpret:
- α (Alpha): Risk-adjusted return not explained by factors — the "skill" component
- β₁ (Market Beta): Sensitivity to market direction
- β₂ (SMB loading): Small-cap vs. large-cap tilt
- β₃ (HML loading): Value (high) vs. growth (low) tilt
- R²: How much of returns are explained by common factors (R² < 0.3 = distinctive strategy)
Step 6 — Walk-Forward Validation
To detect overfitting, split the full history:
- In-Sample (IS): First 60% of data — used to identify strategy parameters
- Out-of-Sample (OOS): Last 40% — held out, tested after IS optimization
Walk-Forward Efficiency Ratio:
WFE = OOS_Sharpe / IS_Sharpe
- WFE > 0.7: Strategy generalizes well — low overfitting concern
- WFE 0.4–0.7: Some degradation — monitor live, reduce leverage
- WFE < 0.4: Likely overfit to historical data — do not trade with confidence
Step 7 — Risk Scenarios
Stress-test the strategy in 3 regimes:
- 2020 COVID Crash: Feb 19 – Mar 23, 2020 (S&P -34% in 33 days)
- 2022 Bear Market: Jan 1 – Dec 31, 2022 (S&P -19.4%)
- 2008 Financial Crisis: Oct 1 – Dec 31, 2008 (S&P -38% in quarter)
For each stress scenario, report the strategy's return vs. SPY return.
Output Format
╔══════════════════════════════════════════════════════════════════╗
║ BACKTEST REPORT — {STRATEGY NAME} ║
║ Period: {start_date} to {end_date} ({N} trading days) ║
╚══════════════════════════════════════════════════════════════════╝
📋 STRATEGY SUMMARY
Universe: {SYMBOL(s) or index}
Entry Signal: {description}
Exit Signal: {description}
Position Type: {Long / Long-Short}
Benchmark: {SPY / QQQ}
Tx Cost: 0.10% per round-trip
📈 CUMULATIVE RETURNS
Strategy: +{X.X}% ($100K → ${value}K)
Benchmark: +{X.X}% ($100K → ${value}K)
Alpha: +{X.X}% annualized
📊 PERFORMANCE SCORECARD
Strategy Benchmark Δ vs. Bench
CAGR: {X.X}% {X.X}% +{X.X}%
Volatility: {X.X}% {X.X}%
Sharpe: {X.XX} {X.XX} +{X.XX}
Sortino: {X.XX} {X.XX}
Max Drawdown: -{X.X}% -{X.X}%
Calmar: {X.XX} {X.XX}
Win Rate: {X.X}% —
Profit Factor: {X.XX} —
Avg Hold: {N} days —
N Trades: {count} —
📉 DRAWDOWN ANALYSIS
Worst Drawdown: -{X}% (from {date} to {date}, {N} days to recover)
2nd Worst Drawdown: -{X}%
Average Drawdown: -{X}%
🧬 FACTOR ATTRIBUTION (Fama-French)
α (Annualized): {X.XX}% [{statistically significant yes/no}]
β Market: {X.XX}
β SMB (size): {X.XX} [{small-cap / large-cap} tilt]
β HML (value): {X.XX} [{value / growth} tilt]
R² (explained): {X}%
🔁 WALK-FORWARD VALIDATION
In-Sample ({dates}): Sharpe {X.XX}, CAGR {X.X}%
Out-of-Sample ({dates}): Sharpe {X.XX}, CAGR {X.X}%
WF Efficiency Ratio: {X.XX} → {Generalizes well / Some degradation / Likely overfit}
💥 STRESS TEST
2020 COVID Crash: Strategy {+/-X}% vs SPY -{X}%
2022 Bear Market: Strategy {+/-X}% vs SPY -{X}%
2008 Q4 Crash: Strategy {+/-X}% vs SPY -{X}%
🎯 VERDICT
Signal Quality: {Strong / Moderate / Weak / Overfit}
Trade Live? {Yes / Caution / No — further development needed}
Suggested Refinement: {1–2 concrete suggestions to improve the strategy}
Limitations
- Backtests suffer from survivorship bias (S&P 500 historical data excludes delisted companies).
- Look-ahead bias is explicitly prevented by using prior-day signals; verify signal construction.
- Slippage is not modeled (real-world execution will have impact beyond transaction cost assumption).
- Fama-French data may have a 1–2 month lag; factors are smoothed — not for very recent periods.
- Past performance in backtests does not guarantee future results.
