Williams %R Mean Reversion: 240 Settings Tested

A Williams %R fade won often and still lost money. Across the 120 long-side settings on hourly USD/JPY bars in 2025 the median win rate was 59.0%, and only 10 of the 120 finished positive. The textbook setup, period 14 buying below −80 and closing above −20, won 68.42% of its trades in 2024 and lost 569.3 pips.

Backtest this setup

Williams %R places the current close inside the high-low range of the last N bars, on a scale from 0 to −100. Below −80 is called oversold, above −20 overbought. The test covers six periods × five entry levels × four exit levels × two directions, 240 settings, on three timeframes over 2024 and 2025.

the textbook 14, −80 / −20highest win rate, 67.0%one dot per setting, 120 of them — 10 finished positive

Each dot is one setting: win rate across, net pips up. The win rates spread from 49% to 67% and almost the entire cloud sits under the zero line.

Why a 68% win rate still loses

Because the average loss is larger than the average win. These are USD/JPY trades that buy when %R falls below the entry level and close when it rises above the exit level, at a 0.3-pip spread.

average winaverage losssecond line: win rate · net pips

The four-hour 79.41% win rate produced only 423.5 pips, because the losing fifth of trades averaged 3.1 times the size of the winners.

The exit rule creates this payoff distribution. A long trade enters below −80 and closes only after %R reaches −20, so a bounce closes the gain while a continued decline leaves the loss open. The result is frequent small wins and occasional large losses.

The high win rate follows from those frequent small bounces; it does not establish positive annual net.

None of the 120 hourly settings in 2025 reached a 70% win rate. The maximum was 67.05%, and that setting still lost 953.7 pips.

Adding Williams %R to a chart

How this was measured

The rule here is directional: "below −80" can only be a buy condition and "above −20" can only be a sell. The long fade and short fade were therefore tested as separate trading rules.

Do the long and short sides differ?

The short side was wiped out in 2024: zero of 120 on both hourly and four-hour bars. USD/JPY rose 1,632 pips that year. Selling "overbought" through a rising market produces exactly this. A counter-trend indicator cannot be judged apart from the direction the market took.

Only the four-hour long side worked: 89 of 120 in 2025, 71 in 2024, and 40 profitable in both. It is the one usable cell in the whole test.

On the four-hour long side, 40 of 120 settings were profitable in both years. Many were profitable in only one year, so one year's result did not identify the next year's period and levels.

Which timeframe actually worked?

Long periods produced the larger results in 2025, while short periods produced the larger results in 2024. Only period 14 was positive in both years (+203 and +197).

Deeper entry levels helped in 2025 (−70 gave +135 pips, −95 gave +757) and made almost no difference in 2024 (+143 and +71). Exit levels showed no consistent pattern either year.

Does a stop fix it?

If the losses are the problem, cutting them is the obvious fix, so stops and targets were measured too.

Every one lost. The four stop-and-target pairs also lost more than doing nothing, in both years. The 24-bar time exit is the single exception, and only in one direction: −989.6 pips in 2025 against −1,082.8 without it, then −729.6 against −569.3 in 2024.

The average loss does shrink, from −110.8 pips to −30.0. But the win rate collapses with it, 68.42% down to 37.04%, and the trade-off goes the wrong way. A stop converted one large loss into a run of small ones.

The size of the losses comes from waiting for an exit condition that may not arrive. Changing the stop distance does not change that design.

Is the spread to blame?

At a zero spread it still loses 1,038.4 pips. This is not a cost problem. At 148 trades a year the spread changes the total by only 296 pips between 0 and 2.0 pips of spread. The 2024 default averaged +44.9 pips per win and −110.8 per loss, so removing trading cost does not repair the payoff imbalance.

Can the four-hour long side be narrowed?

Narrowing the one cell that worked:

The session filters improve 2025 and take 2024 to roughly zero. The only variant positive in both years is the unfiltered period 14, based on 15 to 36 trades a year, which is thin evidence in itself.

These numbers support using Williams %R as one input after direction has been decided elsewhere, rather than letting it choose direction by itself.

Notes

• One pair, two years
• Williams %R varies little between implementations, though conventions differ on whether level comparisons are inclusive. These figures come from Formiq's
• Entries and exits both fill at bar closes
• Spread is fixed at 0.3 pips throughout and slippage is set to zero
• Four-hour settings trade 9 to 92 times a year: thin for reading win rates