Overfitting & Curve Fitting

Overfitting & Curve Fitting

A strategy that looks perfect on historical data is not necessarily a good strategy. It might be a strategy that memorized the past instead of learning from it. This is called overfitting -- and it is the most common reason backtests produce results that never repeat in live trading.

This article explains what overfitting is, why it happens, how to recognize it, and what tools exist to guard against it.

What Is Overfitting?

Overfitting happens when a strategy's parameters are adjusted until they fit the specific price movements in your historical data. The strategy is not capturing a repeatable pattern in the market -- it is memorizing the particular sequence of prices it was tested on.

Think of it like studying for an exam by memorizing the answer key instead of understanding the material. You will score 100% on that exact test. But give you a different version with different questions, and you will fail. The strategy learned the answers, not the concepts.

The middle ground is what you want: a strategy that captures the general trend or pattern without chasing every random fluctuation. A strategy that is too simple (underfit) misses real patterns. A strategy that is too complex (overfit) fits noise that will not repeat.

Why Overfitting Happens

Too Many Parameters

Every parameter in your strategy is a knob you can turn. A strategy with one moving average period has one knob. Add a second indicator with its own period, and you have two. Add a threshold, a filter, and a trailing stop with their own values, and you have five or six knobs.

The more knobs you have, the easier it is to twist them until the strategy fits the historical data perfectly. With enough parameters, you can make almost any strategy look good on any dataset. But that performance is an illusion -- it exists only because the parameters were molded to match that specific data.

Too Little Data

Overfitting is easier when the dataset is small. With 20 trades, a few parameter tweaks can make the strategy win on most of them. With 200 trades, fitting every single outcome is much harder. This is why trade count matters -- more data makes it harder for the strategy to succeed by memorization alone.

Repeated Optimization

Every time you run a backtest, look at the results, change something, and run it again, you are using the results to inform the next change. After 50 iterations, you have effectively tested 50 parameter combinations on the same data. The final version is not independent of the data -- it was shaped by it. Even if you did not use a formal optimizer, manual tweaking over many iterations has the same effect.

How to Detect Overfitting

Parameter Sensitivity

Change one parameter by a small amount -- say, moving a lookback period from 14 to 15. If the results change dramatically, the strategy is fragile. Its performance depends on hitting an exact parameter value, which means it is fitted to the data rather than capturing a broad pattern. A robust strategy should degrade gradually as parameters shift, not collapse.

Results Too Good for the Complexity

A strategy with six indicators and custom thresholds that returns 200% with a 5% drawdown is almost certainly overfit. Simple strategies with a genuine edge produce good results. Complex strategies with amazing results are usually fitting noise. Compare the complexity of the strategy (number of conditions, number of parameters) against the quality of the results. The more complex the strategy, the more suspicious you should be of strong results.

Performance on New Data

The definitive test is whether the strategy works on data it was not designed on. If a strategy performs well on the data you tested and optimized it on but falls apart on a different time period, it was overfit to the original data. In-sample vs out-of-sample testing formalizes this by splitting your data into a design set and a validation set.

How to Avoid Overfitting

Keep It Simple

Use the fewest conditions and parameters that capture the idea you are testing. A strategy with one or two conditions and fixed parameters has limited room to overfit. Every additional condition or tunable parameter increases the risk. If you cannot explain why a condition is there in plain English, it is probably adding noise, not signal.

Use More Data

Test over longer date ranges and shorter timeframes to maximize the number of trades. The more data the strategy has to fit, the harder it is to overfit by accident. A strategy that works across 5 years of data and 300 trades is much less likely to be overfit than one that works across 6 months and 15 trades.

Test on Unseen Data

Split your data. Design the strategy on one portion (in-sample) and validate it on another (out-of-sample). If the strategy performs similarly on both, it is more likely capturing a real pattern. Walk-forward analysis extends this by repeating the split across multiple time windows, giving you a much clearer picture of whether the strategy generalizes.

Overfitting vs. Curve Fitting

These terms are often used interchangeably, and for practical purposes they describe the same problem. If there is a distinction, it is subtle: overfitting refers to the model being too complex for the data, while curve fitting refers to adjusting parameters until the strategy matches historical prices. Both produce the same result -- a strategy that looks great on past data and fails on new data.

Where to Go Next

In-Sample vs Out-of-Sample Testing -- the most direct defense against overfitting.

Walk-Forward Analysis -- repeating the in/out-of-sample split across multiple windows.

Statistical Significance & Trade Count -- why more trades make overfitting harder.

Parameter Optimization -- systematic parameter search and the validation it requires.