Research · Methodology

XAUUSD backtesting methodology: how the model was tested.

The data, costs, rules, checks and limits behind every Tradedge Pulse figure, in one place. If a number appears on this site, this page explains where it came from.

Data & costs

What the backtest was run on

One instrument, one broker data feed, 4.2 years, with costs charged on every trade.

All results are measured in R, the amount risked on a trade. +1R means the trade made what it risked; at 1% risk per trade, +1R is roughly +1% of the account.

InstrumentXAUUSD (spot gold CFD)
DataBroker M15 candles, July 2022 – September 2026, about 98,000 candles
TimeframesZones from the 1H, 2H and 3H charts; confirmation and entries on M15
CostsThe broker's recorded spread plus a slippage allowance on entry and exit, on every trade
PositionsOne trade at a time for the core results; the optional Free Add is reported separately
Result146 trades, +91R, 43% win rate, profit factor 2.71, max drawdown 6.3R (hypothetical)
The rules

What the model does, in concept

Every step is mechanical: two people, or two computers, see the same trades. The exact parameters are private, so the model can't be copied, but the logic is public.

Exits: half the position at the opposite higher-timeframe zone with the rest trailed on structure, the stop moved to breakeven once a trade reaches 1R, and a time exit after five days. The stop always sits beyond the swing and the zone; the median stop was about $23 of gold price. More in our XAUUSD trading strategy.

  1. ContextHigher-timeframe imbalance zones (fair value gaps) are mapped automatically and kept up to date as new ones form and old ones are invalidated.
  2. ReactionPrice must show, on the higher timeframe, that a zone is being defended. Touching it is not enough.
  3. ConfirmationA market structure shift on M15 must confirm the change of direction before any order is placed.
  4. ExecutionA limit entry with a predefined stop. Orders that don't fill in time are cancelled.
Robustness

How we tried to break it

A backtest is easy to overfit. These checks were designed to catch that before any result was published.

TestWhat was doneOutcome
Idea screeningSeveral popular concepts tested before this model, including volatility-based "wall" levelsMost were rejected: they did no better than random levels.
Variation testingHundreds of variants of every rule tested side by sideMore than half of all variants were profitable. The edge is broad, not one lucky setting.
Look-ahead auditEvery decision rebuilt using only information available at that momentNo future data in any signal.
Out-of-sample checkRules settled on 2022–24, then checked on 2025–26Profitable in both: about +33R in 2022–24 and +58R in 2025–26.
Parameter robustnessEach setting shifted one step in both directionsAll neighbouring settings stayed profitable.
Random-entry test400 runs of random trades with identical risk and targetsRandom average about +7R. None of the 400 runs matched the model.
Exit and stop research296 exit plans and 10 stop placements comparedChosen for consistency and low drawdown, not the single highest number.
Platform matchIndicator output compared with the backtest on live-period dataThe same trades, with small differences from broker price feeds.
Automated EAMT5 Strategy Tester, real ticks and 1-minute OHLC, $10,000 at 1% riskProfit factor 3.05 (real ticks) and 3.13, max equity drawdown 4.4%.

What didn't survive is just as useful: see which ICT concepts held up in testing, why stacking 1H, 2H and 3H mattered and what early partial profits cost.

Monte Carlo

What to expect at different risk levels

The backtest trades were reshuffled into 20,000 random orders. The second scenario assumes the live edge is only half of the backtest, which is the one we plan around.

Above 2% risk, the growth figures depend on the edge holding exactly as tested, while drawdowns get deep enough that most traders stop before the recovery. See 1% vs 10% risk per trade.

Risk per tradeBacktest edge: typical growth1-in-20 worst drawdownHalf edge: typical growth1-in-20 worst drawdown
1%2.4×10%1.6×15%
2%5.7×19%2.6×29%
5%56×41%8.4×59%
10%1,227×67%31×86%
Live tracking

The only true out-of-sample test

Since the rules were fixed, every live signal has been posted automatically by the model's MT5 Expert Advisor and recorded on the live results page: new orders, fills, breakevens, partial profits, exits and cancellations, losses included.

Live results include real fills and costs, so they will differ from the backtest. Comparing the two over time is how you should judge the model, and how we do.

Where the numbers are
Limitations

What this testing can't tell you

Hypothetical results have real limits. These are the ones that matter most here.

  • It's a backtest. The results were produced with hindsight and no money at risk. They are not live account results.
  • Same period for development and measurement. The out-of-sample check helps, but the rules were refined on data from the tested period, so live results are likely to be weaker.
  • One instrument, one feed. Other brokers' prices, spreads and swap differ, and news-time slippage can be larger than the allowance used.
  • A modest sample. 146 trades over four years is enough to see an edge but not to rule out a weaker one. A few large winners made a large share of the profit.
  • Markets change. Gold's daily range in 2026 was about twice 2023's in percentage terms. Stops, lot sizes and trade frequency shift with it.
  • Losing streaks happen. The longest in the backtest was 7 trades, and live streaks can be longer.
5 questions

Methodology FAQ

Questions about a specific figure? Email support@tradedgepulse.com.

Are the Tradedge Pulse results real trading results?

The four-year figures are a hypothetical backtest on historical broker data with costs included, cross-checked in the MT5 Strategy Tester. They are not live account results. Live signals are recorded separately on the live results page.

What data was the XAUUSD backtest run on?

Broker M15 data for XAUUSD from July 2022 to September 2026: about 98,000 fifteen-minute candles, or 4.2 years. The higher-timeframe zones (1H, 2H and 3H) are built from the same data.

How were trading costs handled?

Every trade paid the broker's recorded spread plus a slippage allowance on both entry and exit. Results are shown in R, the amount risked per trade, after those costs.

Was the strategy tested out of sample?

Yes. The rules were settled on 2022–2024 data and then checked on 2025–2026, which they had not been fitted to. Both periods were profitable: about +33R and +58R.

How much worse should I expect live results to be?

Plan for roughly half the backtest edge. At 1% risk, the Monte Carlo with half the edge gave about 1.6× growth over the period, with a 1-in-20 worst drawdown of about 15%.