Strategy graveyard / ICT order blocks and the ICT model

Verdict: doesn't work Backtest only

I tested ICT order blocks and the ICT model on up to 10 years of data. Random levels did just as well.

Backtested on SPX (68 sessions), SPY and QQQ (about 3.5 years each) and gold (10 years). Updated 8 Oct 2026.

What are ICT order blocks and the ICT model?

ICT (Inner Circle Trader) is one of the most popular trading schools online. Two of its core ideas:

• Order blocks. Before a strong move that breaks the market's structure, there's usually one last candle going the other way. ICT says that's where large institutions placed their orders, so when price comes back to that candle it should react.
• The "standard model". Price first runs through an obvious level where stop orders sit, like the previous day's high, to grab that liquidity. It then reverses hard enough to break the most recent swing in the other direction (a market structure shift), leaving a fair value gap behind. You enter when price comes back into the gap, with your stop just beyond the sweep. The "Silver Bullet" is the same idea restricted to 10:00-11:00 New York time.

The exact rules I tested

Test 1: order blocks on SPX

1. All four textbook parts were required: a displacement move of at least 6, 8 or 12 points within 3, 5 or 10 one-minute candles; that move had to break the last confirmed swing (a fractal of 5 or 10 candles on each side); the block is the last opposite-coloured candle before the move; and the move had to leave a fair value gap.
2. The zone was either the candle's body or its full range (both tested). The 09:30 opening candle was never used, because SPX's opening print is unreliable.
3. Entry: a resting limit order at the near edge of the zone, first touch only, within 30, 60 or 120 minutes. A close through the far side of the zone before any touch cancelled it.
4. Exit: fixed target and stop pairs in index points (6/6, 10/6, 6/10, 12/8, 10/10), flat at 15:55, with 0.5 points of trading cost per round trip.

Test 2: the ICT standard model on SPY, QQQ and gold

These rules were written down before I looked at any results.

1. Liquidity: the previous day's high and low, plus the overnight high and low where the data had them. A high only counts if the day opened below it (and the reverse for a low).
2. Sweep: a candle trades beyond one of those levels during the trading window.
3. Market structure shift: within 30 minutes of the sweep's extreme, a candle closes beyond the most recent confirmed swing in the other direction.
4. Fair value gap: the largest gap left in that reversal, at least a minimum size relative to recent volatility. No gap, no trade.
5. Entry: a limit order at the gap's edge, or at its midpoint (both tested), before the window closes. Cancelled if price reaches the 2R target first.
6. Stop and target: stop just beyond the sweep's extreme. Targets of 2R, 3R, the nearest opposite liquidity level, or end of day.
7. Variations: 1-minute or 5-minute candles; the full morning (09:30-11:00) or the Silver Bullet window (10:00-11:00); with or without a daily trend filter (only longs above the 20-day average, only shorts below it). One trade a day. Trading costs included.
8. The fake-level check: the same rules were run 20 more times with every level moved by a random amount nearby (a quarter to one full average daily range). If the real levels matter, they should beat most of those fake versions.

Results

Order blocks: blocks did get revisited (54-63% were touched), but what happened next was a coin flip: the target and the stop were hit about equally often. The single best combination looked superb, +3 points a trade with an 85% win rate, but 24 of its 33 trades came in June, the most volatile month. When I scaled the size requirement to each day's volatility so that every month got a fair number of blocks, the result turned negative at every setting.

The ICT model: how often the real ICT levels beat the median fake version, across all the settings tested. 50% is what you'd expect if the levels meant nothing.

On SPY and QQQ the real levels did worse than random ones more often than not. Gold looked better, and its single best setting was positive, but more than half of that came from this year alone. Without 2026 it earned the same as the fake levels. A follow-up test of 84 different stop and target combinations on gold reached the same answer: the fake levels earned about as much as the real ones.

Why it failed

1. The levels aren't special. This is the result that matters. Moving the ICT levels to random nearby prices didn't make the strategy worse. Whatever profit appeared came from ordinary price movement around any level, not from the ICT ones.
2. Size thresholds measure volatility, not quality. A fixed "strong move" requirement mostly fires in volatile markets. The best-looking order-block settings were really just "trade more in June". Once volatility was accounted for, the edge disappeared.
3. Tight stops get hit by noise. A stop placed just beyond the sweep is close by design. On one-minute candles, ordinary noise reached it so often that the win rate was only 22-37%.
4. There are many ways to read ICT. With so many choices (timeframe, window, entry point, target, bias), some combinations will always look good after the fact. That's why the fake-level check matters more than any single result.

Did I test it wrong?

Quite possibly. I tested the version of ICT order blocks and the ICT model described above, with settings I thought were fair, and it didn't work. That doesn't prove every version fails. I may have picked the wrong settings, the wrong market or timeframe, or missed a filter that experienced traders take for granted. Many strategies also depend on judgement that a set of fixed rules can't capture.

If you trade ICT order blocks and the ICT model and it works for you, I'd genuinely like to hear how. Tell me the chart and timeframe, what counts as a setup, any filters, where you get in, where your stop and target go, and when you get out. I'll test your version on the same data, the same way, and publish the result here, whichever way it comes out.

Send me your version

What would change my mind

A precisely defined ICT setup, with every rule written down in advance, that beats the same rules on randomly shifted levels across several years and more than one market. If your version passes that test, I'd like to run it.

Questions

How it was tested

• Order blocks: 68 sessions of SPX one-minute candles from June to September 2026 (26,470 candles), 540 combinations of settings, 0.5 points of cost per round trip.
• ICT model: SPY and QQQ one-minute data from April 2023 to October 2026 (about 850 days each); gold from October 2016 to September 2026 (2,486 days); plus a shorter recent sample on SPY, QQQ, GLD and SLV. Trading costs included. Each setting was compared against 20 fake-level versions.
• Backtest, not live trading: these results come from historical prices with trading costs included. Backtests usually flatter a strategy, so a failed backtest is a strong "no".

This page records my own testing. It isn't financial advice.

Disclosures · Privacy · [email protected]