How we got to the 1‑hour London reversal.

A small, honest edge in one hour of the day.

The 1-hour London reversal is a mean-reversion strategy. On the 1-hour chart of 26 currency pairs, it waits for price to stretch beyond a volatility band around its average, then enters when price snaps back across the Kijun-sen, but only on candles that open between 08:00 and 12:00 London time. The stop sits just beyond the extreme; the target is the opposite band.

Over 22 months of OANDA data with spreads included, holding one trade at a time across all pairs:

The settings were chosen on data up to 6 February 2026. On the eight months after that, which the settings never saw, the average was +0.18R per trade. These are backtest results; see the limits and disclaimer below.

It started with a +330% backtest.

Our starting point was a popular protected indicator built on daily, weekly and monthly pivot levels. Its own backtest on 15-minute Apple stock showed 1,757 trades, a 78% win rate and a +330% return in 19 months, with a 6.8% maximum drawdown.

We rebuilt its logic from the outside. The secret was lookahead: it used the current day's, week's and month's pivot before that period had finished, so on historical candles it was reading the future. With the leak switched on, our rebuild reproduced the profile (+197%, profit factor 5.5). With the leak switched off, the same rules lost 67% with a 28% win rate.

Most ideas died in honest testing.

We held every idea to the same standard: tune on older data, judge on newer data the settings never saw, and include realistic costs. Here is what did not survive.

The 1-hour chart, the London morning.

The Kijun Bands idea came from a swing-reversal approach popular with discretionary traders, built around the Ichimoku Kijun-sen: a stretched move that loses momentum often swings back to the other side of its range. On 4-hour and 15-minute charts that idea failed. On the 1-hour chart, one block of hours stood out: signal candles that open in the London morning.

We tested 144 nearby combinations of settings (session, band width, expiry, exits, buffers). 125 of 144 were profitable on unseen data, which is what a real effect looks like: it does not depend on one magic number.

Later we made the test stricter. Our first version checked the target against the band of the same candle, which a real order could not know until the candle closed. Using the previous candle's band, as a resting limit order works, cut the edge from +0.105R to +0.078R per trade. All results in this paper use the stricter version.

Four rules, in order.

1. Stretch. A 1-hour candle closes beyond the inner band: the 34-period EMA plus or minus 2.5 times the 14-period ATR. The scanner starts tracking the extreme of the move.
2. Snap. The first candle that closes back across the Kijun-sen (26), within 15 candles of the extreme, is the signal. A stretch above the band gives a sell; a stretch below gives a buy.
3. Window. The signal candle must open between 08:00 and 12:00 London time. No new trades on the first Friday of the month (the US jobs report).
4. Plan. Enter at the close. Stop 0.1 ATR beyond the extreme; target the opposite inner band, updated every hour. Hold one trade at a time across all 26 pairs; when two pairs signal together, take the one with the lower spread.

Nothing repaints: every decision uses closed 1-hour candles only. The same logic runs on the AtomicTrades scanner, in a TradingView indicator and strategy, and in a MetaTrader 4 expert advisor, and we checked that they produce the same signals.

The numbers, with costs.

OANDA 1-hour candles for 26 currency pairs, January 2025 to October 2026, with per-pair spread and commission of 1.2 to 3.5 pips per round trip. Settings fixed on data to 6 February 2026; everything after is unseen.

Winners averaged about +0.74R and losers about −1.05R, so the strategy depends on winning roughly two trades in three. When several pairs signalled together and we picked one at random instead of the lowest spread (30 runs), the one-at-a-time average ranged from +0.11R to +0.20R and stayed positive on unseen data in every run.

One block of hours carries the edge.

Taken at every hour of the day, 4,065 signals averaged −0.035R. Grouped by the London hour in which the signal candle closed, the hours from 09:00 to 13:00 were positive in both halves of the data, while the night and the US afternoon lost.

Bar height shows the size of the average result; lime hours were positive and are the ones traded, coral hours lost, grey hours were small and positive but not consistent. One reading: the London morning is when the day's first real liquidity arrives, stretched overnight moves get tested, and the snap back tends to carry. The window was chosen partly after seeing these results, which is why we keep watching it on new data.

The rule that doubled the edge.

Signals cluster: on a busy morning, several pairs that share the US dollar or the euro stretch and snap together. Taking all of them stacks the same bet. Allowing only one open trade across all 26 pairs, and picking the lowest-spread pair when several fire together, changed the profile more than any exit or filter:

Size for the bad run, not the average.

We stress-tested sizing with 2,000 simulated histories that reshuffle whole trading days, so clusters of losses stay together. The headline is about drawdowns, because that is what ends trading accounts:

If the live edge turned out to be only half the backtest, drawdowns at 2% risk rose to about 20%, and at 3% to about 29%. We would not size above 2% until a forward test confirms the edge.

What we would never do

• Martingale (doubling after a loss): in the same simulations, 46% of accounts lost 80% or more. Runs of nine losses in a row happen.
• Full Kelly sizing (about 9% per trade here): the median account lost 81%, because the formula ignores how losses cluster.
• A fixed 5% on every signal, taking every trade: a 13% chance of losing 80% of the account.

We also simulated a two-step evaluation of the kind prop firms run (+10% then +5%, with a 5% daily and a 10% total loss limit). Taking every signal at 1% risk passed only 41% of the time, mostly by hitting the daily limit on busy mornings. With one trade at a time, the daily-limit problem largely disappeared.

What we do not know yet.

• Short history. 22 months of 1-hour data. That is several hundred trades, not decades.
• The window was chosen partly after looking. The London-morning hours were identified from hourly results, which flatters the numbers slightly. Every week of new data is a fresh test.
• Data sources differ. The backtest uses OANDA candles. The free live feed on the scanner uses Yahoo Finance, which can differ by a few pips and occasionally produce a different signal.
• Execution. Spreads widen around the London open and slippage is not modelled. Swaps for trades held overnight are not included.
• Regimes change. Mean-reversion edges can fade. We would reduce risk or pause if the live average stayed below about +0.05R after 60 trades.

The terms, plainly.

Kijun-sen

The midpoint of the highest high and lowest low of the last 26 candles. A slow "fair value" line from the Ichimoku system.

EMA

Exponential moving average; here over 34 candles. The centre of the inner band.

ATR

Average true range: the typical size of a candle, here over 14 candles. Used to scale the bands and the stop buffer.

Inner band

EMA 34 ± 2.5 ATR. A close beyond it is a stretch.

Outer band

Bollinger band: 34-candle average ± 2.5 standard deviations. Shown for context; not required by the rules.

Mean reversion

Trading on the tendency of a stretched price to move back toward its average.

R

The amount risked on a trade (entry to stop). Results in R are independent of account size.

Lookahead

A backtest that uses information not available at the time of the trade. It makes results look far better than reality.

Repainting

An indicator whose past signals change or disappear after the fact, so its history looks better than it traded.

Unseen data

Data that was not used to choose any setting. The fairest test a backtest can offer.

Profit factor

Total gains divided by total losses. Above 1 is profitable.

Research, not advice.

AtomicTrades is a research tool. Nothing in this paper or on the scanner is financial advice, a recommendation to trade, or an offer of any service. AtomicTrades places no orders.

Hypothetical performance results have many inherent limitations. They are prepared with the benefit of hindsight, do not involve financial risk, and cannot account for all the factors of real trading, such as the ability to hold through losses, liquidity, slippage and changing spreads. No representation is made that any account will or is likely to achieve results similar to those shown. Past or simulated performance does not guarantee future results. Leveraged forex trading carries a high risk of loss and is not suitable for everyone.