What Is Curve Fitting? How to Spot an Overfit Strategy

Curve fitting, also called overfitting, is tuning a strategy's parameters until it fits the historical data it was tested on, including that data's noise. The result describes the past precisely and predicts nothing. It is the most common reason a system that looked excellent in testing loses money the month it goes live.

How it works

Every adjustable input is a degree of freedom. Search a grid of values and you are not testing one strategy, you are running one experiment per combination:

Experiments = v1 × v2 × ... × vk

4 parameters × 15 values each = 50,625 backtests
The best of 50,625 is chosen partly for its edge and
partly for the noise it happened to land on

The more experiments you run against a fixed history, the more certain it becomes that the winner owes its ranking to luck. The tester does not report this. It reports the winner as if it were the only candidate.

Four signs, in rough order of reliability:

• Parameter cliffs. Move a setting one notch either way and performance collapses. A real edge degrades gently; a fitted one falls off a table. A moving average that only works at period 47 is the textbook case.
• Too many parameters for the sample. A rule set with six inputs validated on 150 trades is under-determined. There is no fixed ratio, but the direction is not in dispute.
• A suspiciously smooth equity curve. Real returns are jagged. A near-straight diagonal in a backtest usually means either a fitted parameter set or a system that survives by not closing losers.
• Out-of-sample collapse. Performance that holds in the fitting window and evaporates outside it. This is the definitive test, and it is only definitive the first time you look at the out-of-sample data.

Why it matters

A curve-fitted system fails in a specific and expensive way: it does not lose slowly from the start. It behaves acceptably while the market resembles the fitting period, then breaks when the regime changes, usually with size on. The trader reads it as bad luck rather than as the model doing exactly what it was built to do.

The antidote is not a cleverer test. It is a live record, carrying real spread, commission and slippage, long enough to have met conditions the developer never saw.

What the data shows

Across the public accounts on ShowMyTrades with trading history (August 2026): median profit factor 1.28, median deepest drawdown 9.7%, and a median of 171 closed trades per account.

Those three numbers are the honest reference distribution. 38.2% of accounts have reached a drawdown over 20% and 17.6% over 50%. When a strategy is advertised with a profit factor above 3.0 and a maximum drawdown of 4%, it is not claiming talent, it is claiming a position outside a distribution built from 15.4 million synchronised trades. That claim can be true. It requires evidence, and the evidence is live.

The median of 171 closed trades is worth holding onto separately: it is a small sample for validating anything, and it is the sample most live accounts are actually judged on.

Where you see it on ShowMyTrades

The Growth by Trade chart in the charts viewer indexes the curve by trade number instead of by date, which strips out the flattering effect of long idle periods and exposes whether the account is still working or merely still open. The Monthly Returns table shows every red month and every "No Activity" gap — a strategy quietly re-optimised after a bad month usually leaves a visible pause there.

In the Advanced Statistics block, read Profit Factor beside Standard Deviation and Z-Score (Probability): a strong profit factor resting on unusually uniform trade returns describes an account too smooth to have met varied conditions, which is what a parameter set shaped around a single period tends to produce. The Custom Analysis panel's Magic Numbers and Comment Filter let you isolate one expert advisor on a multi-strategy account and read its statistics alone. The Track Record Verified badge, on 65 accounts today, confirms the data came from the broker rather than from a report file.

Common misunderstandings

• "It passed out-of-sample." Once you have looked at the out-of-sample results and adjusted anything, that data is in-sample. You get one look.
• "More history prevents overfitting." What matters is the ratio of observations to free parameters, not the calendar span. Adding years while adding filters can make it worse.
• "Walk-forward optimisation solves it." It reduces the problem and can reproduce it: re-optimising every window on the same instrument is still a search over one history.
• "It's live, so it can't be curve fitted." A system re-tuned every time it stumbles is being fitted in production. The account is real; the edge is still being invented backwards.

To see what separates a record that settles this question from one that cannot, read verified vs unverified track records.

Related terms