Backtest evidence

Does gating entries on ADX or on the slope of a moving average remove losing trades?

Do Trend Filters Remove False Signals? 1,596 Backtests

Finding

No. Six filters were applied to the 38 settings published on this site, across three timeframes and two years. The three ADX levels (20, 25 and 30) lowered the count of settings profitable in both 2024 and 2025 in all nine timeframe-and-level cells: 15-minute 13 down to 6, 4 and 7; hourly 27 down to 18, 10 and 12; 4-hour 16 down to 14, 11 and 12. Comparing what each filter removed with removing the same number of trades at random, 72 of 1,348 comparisons left the 95% range, where chance alone puts about 67 outside. The filtered trades beat the full set on pips per trade in 706 of them, 52.4%.

Key results

ADX filters and two profitable years
Below the unfiltered count in all nine cells. Hourly falls from 27 to between 10 and 18
Against dropping trades at random
72 of 1,348 comparisons outside the range; chance alone puts about 67 outside
Filtered trades beating the full set on pips per trade
706 of 1,348, or 52.4%
What follows a high ADX reading
20.92 pips over ten 15-minute bars against 19.81 for readings under 20; direction continued 48.77% against 50.58%
Which settings a filter helps
The 19 that lost money unfiltered improved in 17 to 19 of 19 in every timeframe and year
Last year's best filter, this year
Beat no filter in 38 of 114 settings
Bars each filter allows
ADX 25 admits 37.58% to 47.24%. The direction filters discard no bars at all

Scope

  • Instrument: USD/JPY
  • Period: 2024-01-01 to 2024-12-31 and 2025-01-01 to 2025-12-31
  • Timeframes: 15-minute, 1-hour, 4-hour
  • Conditions: 38 published settings x 6 filters plus no filter

Method

Of the 43 published article presets, three are excluded: the DMI/ADX article already sets ADX and a rule set holds one ADX configuration, the morning-star and evening-star article joins its conditions with OR so a third condition would widen it, and this article's own preset is a filtered result rather than one of the settings being filtered. The three Bollinger articles publish the same rules, so they count once, leaving 38. Filters are ADX (period 14) at 20, 25 and 30 or more, +DI above -DI, and 50-bar and 200-bar simple moving averages rising. Each is added to the entry rules only; exits stay exactly as each article published them. The control draws the filtered trade count from the unfiltered trades a thousand times and takes the 2.5th and 97.5th percentiles of pips per trade and win rate, from a fixed seed. 38 settings x 7 conditions x 3 timeframes x 2 years = 1,596 runs. Every figure is checked in research/filter-verify.run.ts.

Limitations

  • One pair over two years.
  • Fills are at the close with a fixed spread and no slippage, so live results will differ.
  • Each base setting is one published example per article, not that article's full parameter sweep.
  • Filters are added to entries only; adding them to the exit side was not measured.
  • The 1,348 comparisons are not independent, so the count outside the range is not a significance test.