ICT Silver Bullet (10:00-11:00 New York)

Why we re-ran it

The Silver Bullet is one of the most-asked-about setups, and a first run by one of our agents had already found nothing (t −0.17). We did not publish that run: it used fixed point sizes (a 3-point minimum gap) on an index whose price has risen many times over, took the first gap in either direction with no draw on liquidity, and filled the limit order at the gap on any touch. So we wrote the rules down again, sized everything in average true range, pre-registered them, and ran them once on the 2005-2020 Discovery years.

Making the draw on liquidity mechanical

The draw is the part traders describe with judgement, so it has to be pinned down before a computer can test it. We used the most literal version: the nearest untaken pool among yesterday's high and low and the overnight high and low, fixed at 10:00. Over 4,136 trading days that gave 1,500 orders and 862 filled trades. Most targets were the overnight high or low.

Where a simple backtest goes wrong

Run the same rules the way most backtests are written, on mid prices, filling the limit order whenever price touches it and with no costs, and the result is +0.164R a trade with t 2.74 on 938 trades. That looks like an edge.

Fill details remove about half of it. A limit order at the edge of a gap that price only touches is often not filled in a real market; a backtest that fills it anyway keeps every trade where price turned exactly at the level. And when the order fills inside a minute, nobody knows whether the target was hit before or after the fill, so it cannot be credited. With those rules, stops that gap filled at the open, and still no costs, the result falls to +0.085R a trade, t 1.43.

A small stop makes costs large

The gap's first candle is close to the entry, so the stop is small: a median 5.7 points, from 3.0 points in 2005 to 31.2 points in 2020. A spread and half a point of slippage are then a large share of the risk: up to 0.13R a trade at the median, still up to 0.034R in 2020. After costs the result is −0.050R a trade (t −0.83); charging one more spread plus a point per trade gives t −5.08.

Part of the curve's turn in 2012 is the data, not the market. The minute file shows no spread before 2012 and then 1.00 point against a median stop of 4.3 to 4.5 points: in 2012-2013 the spread alone averaged 0.239R a trade, and the loss of those two years is −53.2R after costs against −27.4R before them. Before costs the rule made +0.17R a trade in 2005-2011 (t 1.83) and +0.02R in 2012-2020 (t 0.23), so the early profit is not significant either.

On 1-minute bars, which many traders use for this setup, the stop is even smaller: the median cost is up to 0.28R and the result is −0.444R a trade (t −10.38).

What else we checked

Five variants were declared before the run and all lose: no draw filter (−0.124R, t −2.73), a draw set by which side was swept before 10:00 (t −1.42), 1-minute gaps (t −10.38), a fixed 2R target (t −1.03) and an entry at the gap's midpoint (t −1.74).

Picking the side of each trade by coin flip instead of by the draw, with the same engine, did as well or better in 15% of 200 runs. Shifting the clock by 1 to 3 minutes either way leaves the result negative at every shift (t −1.31 to −0.31), so the result is not a clock artefact.

What this does and does not say

It does not test the other Silver Bullet hours (03:00-04:00 and 14:00-15:00), other ways of reading the draw, or a trader who skips days that look wrong. It does say that a faithful mechanical version of the 10:00 setup on the Nasdaq-100 has no edge after realistic fills, and that the profit a simple backtest shows comes from how it fills limit orders.