Backtesting & Replay8 min read

Curve Fitting a Trading Strategy: How to Spot It

Replay Trader Editorial Team

Educational content for Indian traders on deliberate practice, chart replay, and setup discovery. Not investment advice.

Curve fitting means tuning a strategy until it explains the past beautifully and fails on unseen days — the pretty equity curve that dies the moment you trade forward or randomise history. It is overfitting dressed as edge: too many parameters, too few honest out-of-sample decisions, too much respect for one heroic NSE bull year.

If you searched curve fitting trading strategy, you may have a backtest that looks deployable and a gut feeling it is fragile. Trust the gut when parameter changes are unexplained, skip rules are missing, and every test day was hand-picked.

Symptoms you can spot without PhD statistics

Parameter archaeology. You tried eleven RSI lengths and kept 13 because 2018–2020 looked best. That is fitting the curve, not discovering stable behaviour.

Missing skip logic. System always in market because flat periods were tuned away. Live trading includes boredom; curve-fit systems do not.

Single-regime sample. Strategy "works" only in trending large-cap years; never tested on range-heavy months.

Discretionary overrides in hindsight. "I would have skipped that" applied after seeing the loss bar — poisons manual backtests too.

Perfection on one symbol. Hero stock dominates results; universe never diversified even for robustness check.

Curve fit vs legitimate refinement

Legitimate refinement Curve fitting
One rule change tied to repeated logged mistake Grid search until equity smooth
Hypothesis stated before test window Best params chosen after seeing all data
Walk-forward or holdout period Full-sample optimisation
Process scores improve on random replay days Win rate only on favourite charts

Refinement responds to process evidence. Curve fit chases past equity shape.

How randomised replay stress-tests discretion

Automated curve fit is famous; discretionary traders curve-fit by practising the same five perfect days until confidence peaks. Randomised NSE historical replay attacks that:

  • Unfamiliar start dates
  • Hidden future forcing real skips
  • Logs tagged by condition — see where rule breaks

If your setup only "works" on days you chose, you fit a narrative, not a market.

Strength of random replay: Cheap out-of-sample feel for discretion.
Failure mode: Small sample anyway if you quit after ten hard days.

Walk-forward thinking (conceptual)

Without quoting fake performance numbers, the idea is simple:

  1. Design rules on period A (or on process logs from replay block A)
  2. Test unchanged rules on period B you did not tune against
  3. Compare process adherence and behaviour, not just green

For discretionary setups, period B can be "next twenty random replay sessions" with frozen card text.

Indian market curve-fit traps

  • Optimising only on low-VIX trending years then deploying into choppy ranges
  • Intraday rules tested without session time filters (9:15–15:30 IST)
  • Ignoring costs on high-turnover coded systems
  • Single-cap hero narratives from memed stocks — not repeatable universes

Use liquid, defined universes for honest tests. Hypothetical examples only; no stock tips.

Strengths and failure modes of anti-curve-fit discipline

Strengths: Saves months live on fragile systems. Encourages simpler rules. Pairs with look-ahead controls for cleaner research.

Failure modes: Paralysis — never deploying even after honest process evidence. Confusing "simple" with "undefined." Using random replay once, getting uncomfortable results, and tweaking rules mid-sample until comfortable again — refitting under another name.

Practical de-fit protocol

  1. Freeze setup card text for two weeks of random replay
  2. Log process scores only — no parameter changes allowed inside window
  3. Review condition tags where scores drop
  4. One change permitted after window; repeat

If after three windows the setup still needs constant rescue, consider drop — negative result worth knowing before live size.

Simplicity as anti-fit discipline

Rules with fewer moving parts survive random replay better because each component gets enough occurrences to judge. A setup with five filters may show a gorgeous backtest simply because filters eliminate losing days post hoc — each filter another degree of freedom eating your sample.

When simplifying, remove filters one at a time across replay windows rather than all at once. Watch process scores, not equity curve cosmetics.

Relationship to look-ahead bias

Curve fitting optimises the past; look-ahead bias steals from the future. Many overfit systems depend on both: parameters tuned on full sample plus signals that peek. Cleaning look-ahead alone does not de-fit; freezing rules on random replay windows addresses both discretionary and narrative fitting.

Where Replay Trader fits

Replay Trader emphasises randomised NSE historical practice with hidden futures — a discretionary stress test curve-fitters often skip. If your equity curve came from hand-picked charts, rebuild sample blind via join the waitlist. Educational simulation; not a guarantee of edge.

Educational note

This article is for educational trading practice only. It is not investment advice, not a recommendation to buy or sell any security, and not a promise of future results. Practice outcomes do not guarantee live trading outcomes.