ICT Silver Bullet (New York AM) on NAS100

Method

The rules were written down before the test and every parameter was pinned: a swing of 4 bars each side, a fair value gap of at least 0.2 ATR within 6 bars of the sweep, a limit entry at the near edge of the gap, the stop 2 points beyond the sweep's extreme, a 2R target, a 2-hour time exit, one trade per day, 1% risk per trade. Nothing was optimised.

Data: NAS100 5-minute bars from a broker's MetaTrader 5 feed, 1 October 2021 to 23 October 2026 (358,580 bars). Signals fill at the next bar's open or at the limit price; if one bar touches both the stop and the target, the stop is assumed to fill first.

The test is a rolling walk-forward with five folds: each fold has a training window of about 19 months followed directly by a test window of about eight months, and only trades inside the test windows count. The first 19 months (October 2021 to May 2023) are training only; the test windows run from May 2023 to October 2026. Because every parameter was fixed before the test, the training windows had nothing to fit: the optimise step ran one trial with the fixed settings and every fold used the same values. Each test window is simply the same rules applied to the next stretch of data. Results are in R, the amount risked per trade, and are before trading costs: no spread, commission, slippage or financing is charged.

Verdict: reject. Not this time. Out of sample, the Silver Bullet written as fixed rules lost 0.117R per trade on NAS100, and all five test windows lost money. All figures are before trading costs.

Run run_d87f236a61, IDFX engine, 23 October 2026.

The idea

The Silver Bullet is a one-hour setup taught by ICT. Between 10:00 and 11:00 New York time, price runs beyond a recent swing high or low (a liquidity sweep). Then it turns sharply the other way and leaves a fair value gap. You enter when price comes back into the gap, put the stop beyond the sweep, and aim for a multiple of the risk.

The rules, exactly as tested

Every number was fixed before the test and none was optimised. The values follow the commonly taught version of the setup; they were not chosen by searching this data.

• Window: signals only between 10:00 and 11:00 New York time.
• Sweep: price trades beyond a swing low (or high), defined as 4 bars on each side, and closes back inside.
• Displacement: within 6 bars (30 minutes) of the sweep, a three-bar fair value gap forms in the opposite direction. It must be at least 0.2 × ATR in size.
• Entry: a limit order at the near edge of the gap: the top of a bullish gap, the bottom of a bearish one. The order is cancelled if it is not filled within 6 bars.
• Stop: 2 points beyond the extreme of the sweep.
• Target: 2 × the risk (2R).
• Time exit: flat after 24 bars (2 hours), and flat before the weekend.
• Frequency and size: at most one trade per day and one position at a time, risking 1% of the account per trade.
• Direction: both. Longs after a sweep of a low, shorts after a sweep of a high.

Data

• NAS100 (Nasdaq 100 CFD), 5-minute bars, from a broker's MetaTrader 5 price feed.
• 1 October 2021 to 23 October 2026: 358,580 bars over 5.06 years. All of them are genuine 5-minute bars, with no coarser history mixed in.

How the test works

Fills. A signal on one bar can only fill on a later bar, at the limit price, or at that bar's open if price opens beyond the limit. If a single 5-minute bar touches both the stop and the target, we count the stop, because a bar does not show which came first.

Walk-forward. The test is a rolling walk-forward with five folds. Each fold has a training window of about 19 months, followed directly by a test window of about eight months. The test windows follow each other without overlap. Only the trades inside the test windows count as the result.

The first 19 months (October 2021 to May 2023) are used only for training and are never tested. The last 3.5 years (May 2023 to October 2026) are the test windows.

Nothing was trained or tuned in this run. In a normal walk-forward, each training window is where the optimiser searches for the best parameters, which are then tested on the window after it. Here every parameter was fixed before the test, so the training windows had nothing to fit. The engine's optimise step ran one trial with the fixed settings and returned them unchanged (optimize.json: method "none", an empty search space), and every fold used the same values. Each test window is just the same rules applied to the next stretch of data. The walk-forward still does two jobs:

• it reports the result window by window, so you can see whether the rules held up over time or only in one lucky stretch;
• it keeps the headline result to data after the first training window.

The out-of-sample result therefore covers 11 May 2023 to 23 October 2026, about 3.5 years, not the full five.

Costs. None are charged: no spread, commission, slippage or overnight financing. This study asks whether the signal itself has an edge. Real trading costs would only make the numbers below worse.

Results (out of sample, before trading costs)

R is the amount risked on a trade. −26.9R means the account lost 26.9 times its per-trade risk; at 1% risk per trade, the account fell from 10,000 to 7,563.

How the trades ended. Only 16 of the 230 trades (7%) reached the 2R target. 96 hit the stop, and 118 were still open after two hours and were closed by the time exit. Most of the "wins" are small gains closed by that time exit, not trades that reached the target.

Window by window. Every test window lost money:

Longs and shorts. Both lost. Longs: 109 trades, −0.075R each, −8.2R in total. Shorts: 121 trades, −0.154R each, −18.7R in total.

The full five years, for context. Because nothing was fitted, the same fixed rules can also be run over the whole period, including the first training window. That run took 344 trades and lost 0.063R per trade, −21.5R in total. The first training window on its own (October 2021 to May 2023) was slightly positive: +5.3R over 114 trades. The rules made a little money in 2021–2022 and lost in every year after. That early stretch is not part of the headline result.

The verdict

Reject. The checks that decide it:

• Positive expectancy: failed. −0.117R per trade. Resampling the trades 5,000 times puts the average between −0.222R and −0.012R in 90% of cases, and above zero in only 3.4% of them.
• Fold consistency: failed. 0 of 5 test windows were profitable.
• Beats shuffled price: failed. The rules were also run on 40 price series built by reshuffling blocks of the real data. Those random series scored as well or better in 34 of the 40 (p = 0.85). The rules do no better than chance on prices with no Silver Bullet structure in them.
• Neighbouring settings: failed. Nudging the parameters a little either way leaves the result negative in 95% of cases. This is not one unlucky setting.
• Sample size: passed. 230 trades is enough to measure the result.

Some of the engine's checks are built for strategies whose parameters are optimised, and they mean little or nothing here because nothing was optimised:

• Probability of overfitting (PBO): not computed. It compares several candidate parameter sets, and this run had only one.
• Deflated Sharpe: the "number of combinations searched" it corrects for is just the same fixed settings counted once per window. The Sharpe ratio is negative anyway, so there is nothing to deflate.
• Walk-forward efficiency (−248%): this normally measures how much of an optimised result survives on new data. The verdict's wording, "the optimizer was fitting noise", does not apply, because there was no optimiser. Here the figure only says that the first window (2021–2023) did better than the later ones.
• Robust to trade order: reshuffling the order of trades cannot change their sum, so "0% of reshuffled sequences end profitable" just restates that the total is negative.

The rejection does not rest on any of these. It rests on the negative out-of-sample expectancy, losses in every window, and a result that random price series match.

Limitations

• Before trading costs. Spread, commission and slippage would make the result worse.
• 3.5 years out of sample, not 5. The first 19 months are the walk-forward's first training window and are not in the headline result. Nothing was fitted in them.
• One precise reading of the Silver Bullet. Other ways of defining a sweep, a fair value gap or the stop might behave differently. This study does not test them, and a different definition is a different strategy that needs its own test.
• One instrument, one session, one price feed: NAS100, 10:00–11:00 New York, one broker's MT5 data.
• 5-minute bars. A bar that touches both the stop and the target is counted as a loss.
• A backtest. It is not live trading or a funded-account result, and it says nothing about either.

Limitations

• Before trading costs. Spread, commission and slippage would make the result worse, not better.
• The out-of-sample result covers about 3.5 years (May 2023 to October 2026), not the whole five; the first 19 months are the walk-forward's first training window, in which nothing was fitted.
• One precise reading of the Silver Bullet. Other definitions of a sweep, a fair value gap or the stop could behave differently; this study does not test them.
• One instrument (NAS100), one session (10:00 to 11:00 New York), one broker's price feed.
• 5-minute bars: what happens inside a bar is not visible, so a bar that touches both stop and target is counted as a loss.
• A backtest, not live or funded-account trading.