Backtest vs forward test: por qué tu sistema funcionó en el pasado y no en el mercado real
Quick answer: a backtest optimises against data you have already seen; a forward test applies finished rules to data the system never saw. It almost always performs worse, and that drop is the valuable information. If your system loses half its edge on fresh data, the original edge was largely a fit to the past, not a signal.
Puntos clave
- The backtest validates the idea; the forward test validates that you did not overfit it.
- The more parameters you tune, the bigger the drop on new data.
- Moderate degradation is normal and expected; 80% is not.
- A live demo adds what no history contains: execution.
The three tests and what each answers
| Test | Question it answers | What it misses |
|---|---|---|
| Backtest | Did the idea ever have an edge? | Overfitting, real costs, execution |
| Historical forward test | Does it survive data it never saw? | Real slippage and your behaviour |
| Live demo | Does it work with real execution? | The pressure of your own money |
All three are necessary and none replaces the next. Skipping the second is the most common reason a “perfect” system stops working in its first month.
How to run an honest forward test
- Split the history before you touch anything. Set aside the most recent 30% and do not look at it.
- Develop and tune only on the first 70% until the rules are completely finished.
- Freeze the parameters. From here on, not a single figure changes.
- Run the reserved 30% once, with real costs included.
- Compare the two expectancies and see how far it fell.
Step 4 is worth exactly one use. If you look at the reserved slice, adjust, and look again, you have just turned it into part of the backtest. From that moment you have no clean data left to validate anything, and you must wait for time to pass to get more.
The full version of this process — how to pick the cut-off date, what “frozen” has to include, how many reserved trades you need and how to read the result — is in the free ebook The Forward Test Handbook, which also includes a printable log sheet and a ten-step checklist.
How much degradation is normal
| Drop in expectancy | Reading | Qué hacer |
|---|---|---|
| Under 20% | Very good sign, robust rules | Move to demo |
| 20 to 40% | Normal and expected | Move to demo with adjusted expectations |
| 40 to 70% | There is overfitting | Simplify: fewer parameters |
| Over 70% | The edge was fitted noise | Start again |
Indicative ranges. What matters is not the exact number but the order of magnitude: a small drop suggests robust rules, an enormous one suggests you fitted the strategy to that period's noise.
Why more parameters make the drop worse
Every parameter you optimise is another chance to fit the period's noise instead of the signal. A system with two rules and one filter usually survives new data reasonably well; one with eight chained conditions almost always collapses, because it described the past rather than the market. If the only way to make your backtest shine is to add one more condition, that condition is the problem.
Mistakes that inflate the result
- Optimising on the whole history and then “validating” on the same data.
- Testing a hundred combinations and keeping the best without penalising the search.
- Leaving spread, commission and swap out of the simulation.
- Using closing prices for entries that would happen intrabar in live trading.
- Choosing the historical period that happens to suit the strategy.
The last step before real money
When the historical forward test holds up, the live demo remains — the only thing that genuinely measures slippage and latency. Two or three weeks is enough to see whether execution matches the simulation. To log both phases in the same format there is the free ebook The Trading Journal Blueprint, and to include costs correctly, The True Cost of Trading. If what you are validating is an automated system, the five-day Automation Test exists for exactly that phase, and Cortex Automation then runs with the same parameters you tested.
What exactly is the difference between the two?
A backtest tunes the rules while looking at the data; a forward test applies finished rules to data that took no part in the tuning.
How much history should I reserve?
Around 30% of the total is usually enough, provided that slice contains at least 100 trades.
Can I skip the forward test if the backtest is excellent?
It is the other way round: the more spectacular the backtest, the more necessary the forward test. Perfect results are usually overfitting.
How long should a live demo run?
Two to four weeks for execution; validating expectancy takes trades, not days.
What if the system stops working later, live?
It happens, which is why it is worth repeating the process periodically on new data rather than trusting a single validation.
Validate on data the system never saw. Start with the free The Forward Test Handbook, then try it with the Automation Test or see Cortex Automation.
Disclaimer: educational content. The ranges are indicative and should be validated with your own data. Cortex Next is not a financial service and guarantees no results. Trading involves risk of loss.
