How curve fitting makes a backtest look great in-sample yet fail live, the warning signs to watch for, and how version traceability makes over-optimization visible.
Start with a no-code crypto spot strategy, lock the version, run the backtest, and keep the result traceable for comparison.
A backtest can show a clean equity curve, a high win rate, and a strong return, and still tell you almost nothing useful. Overfitting in backtesting is the reason. When you tune a strategy until it fits the random noise in one slice of history, the result looks excellent on that data and collapses the moment conditions change. Learning to spot the difference between a good strategy and a lucky one is the single most important research skill, and it is the whole reason a backtest exists.
Traseq is a research workspace, not a live trading or exchange-execution platform. It does not place orders, connect to exchange accounts, or guarantee performance. The point of this post is the opposite of a guarantee: it is to help you distrust results that look too good.
Overfitting, also called curve fitting or over-optimization, happens when a strategy describes the specific past data rather than the underlying market behavior. Every price series contains two things: signal (repeatable structure you might exploit) and noise (random wiggle that will not repeat). The more knobs you turn and the more you tune them to maximize a single backtest, the more your rules end up explaining the noise.
The trap is that overfitting improves your backtest. Adding a filter, nudging a threshold, or excluding a bad month almost always makes the in-sample number go up. That is exactly why a great backtest is not evidence of a good strategy. It can simply be evidence that you tried hard enough to fit the history.
A robust strategy is the opposite: it gives up some in-sample perfection in exchange for behavior that holds up on data it never saw.
You rarely catch overfitting by staring at the return. You catch it by looking at how the result was produced and how fragile it is. Watch for these signs:
Small trade counts are the most common and most underrated trap. The interactive demo on Traseq's Learn hub runs three textbook templates on real BTC/USDT 1h candles from 2024-11-03 to 2024-12-31. The RSI Mean Reversion template finished that window at +1.74% — but it took only 9 trades to get there.
Nine trades is not enough to trust. Reorder one or two of them, or shift the window by a week, and the return could easily flip negative. With so few samples, you could "improve" that result by tuning the RSI thresholds until they happen to catch the winners and avoid the losers in this specific window — and you would learn nothing about whether the rule works. The fewer trades a backtest produces, the more any single result is dominated by chance, and the easier it is to fit the noise.
It helps to see what unflattering, un-tuned results look like. In that same choppy two-month window, all three demo templates were run as-is, with no optimization:
The window was a sideways-to-down chop after a rally, and it shows: two trend/breakout templates net-lost and mean reversion only just broke even. That is not a failure of the demo — it is the honest picture. The danger is that you could "fix" any of these by adding filters until the line points up. That edited result would look great and mean nothing. The discipline is to keep the rule simple, accept the honest number, and treat the backtest as a question, not a verdict.
You cannot eliminate overfitting, but you can make it much harder to fool yourself:
A related warning: a high win rate is one of the easiest numbers to overfit toward, yet it can still lose money. See why a high win rate can still lose money.
Overfitting hides in the gap between versions — the small edits you made to chase a number and then forgot. Traseq is built to close that gap.
Seeing the full chain of edits is what turns "I have a great backtest" into "I can tell whether this strategy was good or just lucky."
The fastest way to build intuition is to watch honest results in motion. Run the no-signup interactive demo and see how a small change in assumptions moves the outcome, then read Core Concepts to understand the execution model behind every run.
Overfitting is when a strategy is tuned to fit the random noise in one slice of historical data instead of repeatable market behavior. It makes the backtest look excellent in-sample but causes the strategy to fail on new data, because the rules were shaped around details that will not repeat.
They describe the same problem. "Curve fitting" emphasizes shaping the rules to trace one specific price path, and "over-optimization" emphasizes tuning too many parameters to maximize a single backtest. Both result in a strategy that fits the past rather than the market.
Look for warning signs rather than the headline return: too many tuned parameters, a suspiciously smooth equity curve, a tiny trade count, and results that collapse when you shift the date range or add realistic fees. If small changes break it, it is probably overfit.
With only a handful of trades, luck dominates the result, so any single number is unreliable. A strategy that nets a positive return on 9 trades could easily flip negative if a couple of trades were reordered. Few trades also make it easy to tune thresholds until they happen to catch the winners in that one window.
Traseq ties every backtest to a finalized strategy version, so the exact parameters behind each run are recorded, and comparison sets let you place versions side by side across performance, risk, conditions, and period. That makes it visible when a result only improves because you added knobs or cherry-picked a date range.
| Sign | Why it points to overfitting |
|---|
| Too many parameters | Each lookback, threshold, stop, and filter is another degree of freedom to fit noise. A rule with eight tuned numbers can fit almost anything. |
| A suspiciously smooth equity curve | Real strategies have losing streaks. An equity curve with almost no drawdown over a noisy market usually means the rules were shaped to dodge specific past losses. |
| Results that collapse when you shift dates or fees | If moving the start date a few weeks, or adding realistic fees and slippage, turns a winner into a loser, the edge was never robust. |
| Tiny trade counts | A handful of trades cannot distinguish skill from luck. A "great" result built on 9 trades is mostly noise. |
| A long history of tweaks chasing one number | If you edited the rules dozens of times and kept whatever raised the backtest return, you optimized the workspace, not the market. |
| Template | Return | Win rate | Trades | Profit factor |
|---|
| SMA(200) Trend Filter | -6.89% | 22.7% | 22 | 0.36 |
| RSI Mean Reversion | +1.74% | 44.4% | 9 | 1.12 |
| Donchian Breakout | -10.27% | 34.5% | 29 | 0.66 |
1h, 4h, 1d) and across different date ranges and market regimes. A robust strategy degrades gracefully; an overfit one falls apart.Apr 24, 2026