In 2025 HSBC and IBM reported that a hybrid quantum model improved their prediction of whether a bond-trading request would be filled by about 34% over their classical baseline. Since then, "quantum" indicators have appeared all over retail trading platforms. Most are ordinary scripts with a new name.
We wanted to know whether the real thing, a quantum model fed with order data, could help forex traders. So we built one and tested it the same way we test everything: train on old data, then judge it only on data it has never seen.
The data
Forex has no central order book, but OANDA publishes snapshots of its clients' pending orders (the order book) and open positions (the position book). We collected:
- 8 pairs: EUR/USD, GBP/USD, USD/JPY, AUD/USD, USD/CAD, USD/CHF, NZD/USD, EUR/JPY
- Two snapshots a day, June 2018 to October 2026
- About 33,700 usable order book and position book pairs, plus hourly prices
The features (the "liquidity pool" signals)
From each snapshot we measured, within 0.6% of price:
- Stop imbalance: buy stops above price vs sell stops below (where the liquidity pools are)
- Limit imbalance: buy limits below vs sell limits above
- Distance to the biggest stop cluster above and below, in daily ranges
- Crowd positioning: net long percentage, change since the last snapshot, and how many traders are trapped in losing positions
- Momentum over 1 day and 5 days
Each feature was turned into a rolling z-score per pair, as the quantum-indicator roadmaps recommend.
The models
- Simple rules on each feature (trade toward or away from it)
- Logistic regression (classical baseline)
- Quantum-kernel classifier (QSVM) with a 6-qubit ZZ feature map, simulated exactly
- Variational quantum classifier (VQC), 6 qubits and 3 layers, the "indicator brain" in most quantum trading guides
Training data: 2018 to 2023. Settings chosen on 2024. Final judgement: January 2025 to October 2026, untouched until the end. Every trade paid a realistic spread.
Results on unseen 2025 to 2026 data
The test: predict which way price would move 0.5 daily ranges first within 48 hours.
| Model | Trades | Net result |
|---|---|---|
| Logistic regression | 918 | -41.4R |
| Quantum kernel (QSVM) | 4,564 | -160.1R |
| Variational quantum classifier | 7,560 | -329.1R |
None made money. The VQC could not learn anything beyond a coin flip (its training loss stayed at 0.693, which is the value for pure guessing).
The best single signal was the stop imbalance: price did lean slightly toward the bigger pool of stops, by about +0.02R to +0.03R per trade in most periods. That is real, but it is smaller than the spread, and over a 5-day horizon it flipped sign in 2024. It is not a trading edge.
What we did find: crowded stops get hit more often
We then asked a different question: does it matter where you put your stop?
For stops placed 1 to 1.5 daily ranges away (swing-trade size), a stop sitting inside a crowded stop zone was hit more often than a stop at the same distance in an empty zone, in every period we tested:
| Stop distance | 2018 to 2021 | 2022 to 2024 | 2025 to 2026 |
|---|---|---|---|
| 1.25 daily ranges | 38.0% vs 33.1% | 40.3% vs 36.5% | 39.1% vs 34.2% |
| 1.5 daily ranges | 29.7% vs 24.1% | 31.2% vs 27.9% | 30.3% vs 26.9% |
(Hit rate within 5 days, crowded zone vs empty zone.)
So the "liquidity pool" idea is real in the data: stops bunched in obvious places do get taken more often. But when we moved crowded stops just past the cluster, the gain was the same as simply widening any stop by the same amount. The cluster itself added no extra edge once the wider stop and smaller position were accounted for.
What this means for traders
- A quantum label does not create an edge. The same features failed in classical and quantum models alike. The bottleneck is the information in the data, not the computer.
- Retail order book data is a weak directional signal. It shows where retail stops are, not where banks will push price.
- Use it for stop placement awareness, not entries. If your stop sits in an obvious cluster, expect it to be tested more often, and size the trade so a slightly wider stop costs the same risk.
What we changed in our engine
Nothing, on purpose. Adding a signal that fails on unseen data would make our results worse, and we only ship what passes. Engine v6 keeps its tested swing rules and AI filter. We will re-run this test when longer order book history or new data sources become available, and publish the result either way.
Read more about the hardware side in our quantum computing bottlenecks case study, and about stop clusters in our order book guide.
Frequently asked questions
Do quantum trading indicators work?
In our test of a real quantum-style classifier on 8 years of order data, no. Most retail "quantum" indicators are classical scripts with a marketing name.
Is OANDA order book data useful?
It is useful for seeing where retail stops and orders bunch up. In our tests it did not predict direction well enough to beat the spread.
Did the HSBC result apply to forex?
The reported HSBC and IBM work was on bond trading requests and fill probability, which is a different problem from predicting currency direction.
Next steps
- See our tested and live record on the Performance page.
- Size your trades with the Risk Lab.
- Get today's free signals.
*Educational content, not financial advice. Trading carries a high risk of loss. Backtested results do not guarantee future results.*
