1nTest · strategy backtester and trading bot builder

Do backtest filters survive a new year? Masks fitted on 2022–2024, checked on 2025

Filters chosen on 2022–2024 (session, spread to stop, stop size, assets to skip) took five strategy templates, which together lost −0.14 R per trade, to +0.58 R. In 2025 the same filters made −0.11 R per trade against −0.13 R without them, but a random pick of the same number of trades made −0.09 R and did at least as well in 54% of draws. The in-sample improvement mostly did not carry over.

What we did

For each of 5 strategy templates we pooled the trades of twenty assets, as one multi-asset backtest. On 2022–2024 we added filters one at a time, each time taking the single filter that improved the result most, until none added more than 1 R. The candidates were the same kinds of filters a trader sets on a backtest page: skip a session, cap the spread as a share of the stop, bound the stop size in ATR, skip an asset.

That produced 48 filters in total and kept 8% of the in-sample trades. Then we applied them, unchanged, to 2025.

The result

Without filters
2022–2024 (fitted)−0.14 R
2025 (new data)−0.13 R
With filters
2022–2024 (fitted)+0.58 R
2025 (new data)−0.11 R
Random trades, same number
2022–2024 (fitted)—
2025 (new data)−0.09 R

Net result per trade after costs, in R. In 2025 the filters kept 129 of 1,808 trades. They improved the per-trade result for 3 of the 5 templates.

Why the comparison is with random trades

When a strategy loses money, any filter that removes trades improves the total, simply because fewer losing trades are taken. The fair test is whether the filter picks better trades than chance. Here a random selection of the same number of trades from each template did as well as the filters in 54% of draws, which is not distinguishable from luck.

What to do instead

  • Fit filters on one period and judge them only on a later one, per trade and against a random selection of the same size.
  • Be most suspicious of filters that skip individual assets or narrow parameter bands: they fit the past most easily.
  • A filter with a reason that does not depend on the sample, such as skipping trades where the spread takes a large share of the stop, is more likely to hold.

How we measured

  • Five strategy templates on twenty assets, engine backtests on the production server, period 2022-01-01 to 2026-01-01, fixed risk of 1% per trade, commission, swap and bid/ask prices.
  • Sessions by the New York clock: Asia 18:00–02:00, London 02:00–08:00, New York 08:00–17:00.
  • Greedy selection on 2022–2024: a filter had to keep at least 20 trades and add more than 1 R to the total.
  • Placebo: for each template, the same number of 2025 trades as the filters kept, drawn at random; repeated 2,000 times.

Backtests run on October 24, 2026.

Questions

So filters are useless?

No, but a filter found by searching the past needs the same test as the strategy itself: data it was not chosen on, and a comparison with chance.

Why did the filtered result stay positive in 2025?

Mostly because the filters removed almost all trades of the worst template. A random selection of the same size per template is also close to zero or above.

For education, not investment advice. Past results, including backtests, do not guarantee future results.

← All research