Trend-Following Backtester · Guide · Concepts 한국어
Concepts
Optimization and overfitting — why backtested returns disappear in live trading
“I found the optimal parameters and it returned 100% a year” is usually true, and usually useless. Those values were chosen to fit a past that has already happened.
What optimization is
Every strategy has numeric settings — how many days in the average, how many standard deviations wide the band. Changing them changes the result, so you search for the combination that performed best on history. That is parameter optimization.
The problem is right there. “Performed best” is a statement about the past. Nothing guarantees the same values are best next year.
Measuring the gap
Rather than argue about it, split the data. Find the optimal parameters using only the first part, then apply them unchanged to a second part the search never saw. That is the live-trading condition.
| Segment | CAGR | Sharpe |
|---|---|---|
| In-sample where parameters were chosen · through 2023-12-05 | 111.9% | 1.92 |
| Out-of-sample unseen · 2023-12-05 onward | 33.9% | 1.26 |
| Buy and hold, same window | 17.2% | 0.59 |
볼린저 밴드 돌파 fell from 111.9% to 33.9% — roughly a factor of 3.3. What vanished is the size of the overfit.
But read the third row before concluding. Buy and hold returned 17.2% over that same unseen window. The strategy still beat holding on data it had never seen — it just did not deliver what the backtest advertised.
Reducing overfit
- Fewer parameters. More knobs, easier to fit noise. The strategies here have two or three.
- Split the data. In-sample/out-of-sample as above, or repeated re-optimization across windows.
- Check the neighbourhood. If the best value is a spike and its neighbours are terrible, it is luck. You want a broad plateau.
- Test across assets. A value that only works on one instrument is probably fitted to that instrument's past.
How to read the numbers on this site
The asset pages here report parameters optimized over the full history. As shown above, that runs higher than live trading would. They are published anyway because with two or three parameters the distortion is bounded, and the relative ranking between strategies remains informative.
Use them to see which strategy suits an asset's character. Do not use them as a return target.