Most backtests look great. That's the problem. They use overfit rules, ignore execution costs, and reward the person who built them instead of the strategy itself.
Why backtest: Validates statistical edge vs bull market luck. Shows behavior across regimes. Reveals worst drawdown. Gives reference point when real trades go against you. Honest caveat: goal isn't a strategy that looks good on paper — it's one that would have been hard to break.
Step 1 — Surgical precision: If you can't explain your entry in one sentence it's not testable. Entry: RSI crosses above 30 while price above 200 EMA. Exit: RSI above 70 OR price drops below 200 EMA. Stop: 2x ATR. Size: 2% risk. Any two people should make identical trades.
Step 2 — Data + regime testing: Min 3-5 years spanning bull and bear. Ideal 10+ years including 2008, 2020, 2022. Critically: segment by regime. A strategy that only works in bull markets is just market beta — you could've bought an index fund.
Step 3 — Run it: CompoundPulse Backtests page — select ticker, timeframe, strategy. No code. Don't adjust parameters based on early results — that's curve fitting.
Step 4 — Metrics + walk-forward: Total return, vs buy & hold (most strategies don't beat it), win rate (40% with 2:1 R/R is profitable), profit factor (>1.5 solid, >2.0 excellent), max drawdown (stress test), Sharpe ratio (>1.0 acceptable, >2.0 strong), number of trades. Then: split into in-sample (optimize) and out-of-sample (holdout). If it only works on data you tuned it for, it found patterns in noise, not signal.
Step 5 — Equity curve: Grows steadily with shallow drawdowns. Red flags: 30%+ drawdowns, long flat periods, only works in bull markets, suspiciously smooth curve (real strategies have rough patches — too smooth = overfitting, not skill).
Common Mistakes (the heart of the article):
Curve fitting: You can make any strategy look great on 5 years of AAPL if you tweak enough parameters. You haven't found an edge — you've memorized the data. Real test: does it work on TSLA? Financials 2018-2022? If it only works when conditions are precisely right, you have a coincidence.
Survivorship bias: Lehman Brothers, Enron, Bed Bath & Beyond aren't in your universe. Biases every result upward. Always in the same direction.
Parameter fragility: Strategy only works with RSI 14 but falls apart at RSI 12 or 16 = fragile. Real edges are robust across parameter ranges. Changing one input by 10% shouldn't destroy results.
Not enough trades: Under 20 is astrology. Need 30-50 minimum. A coin flipped 12 times can "win" 70% without meaning anything.
Slippage: High-frequency strategies on thin stocks look phenomenal and are completely unexecutable. Stock trades 50K shares/day, you need 10K at open — the historical fill price is fiction.
Honest filter: After realistic slippage and commissions, does the strategy still have an edge? If it only works in a frictionless world, it doesn't work.
3 Questions before trusting any backtest:
Does it work on out-of-sample data? (Performance collapses = overfitting found)
Does it work across different assets and sectors? (Robust edges generalize, coincidences don't)
Does it still work if I change parameters slightly? (RSI nudged by 2, MA shifted by 10 days — still consistent = maybe real; falls apart = fragile artifact)
After all 3 pass → paper trade in real time with simulated money before risking capital. CompoundPulse has both built-in — no switching platforms, no code.
If you want to run this workflow without stitching tools together, try CompoundPulse — backtesting, charting, and screening in one place. compoundpulse.com
Top comments (0)