VWAP 'holy grail': long above VWAP, short below (Nasdaq 100, 5-minute)
The 'VWAP holy grail' rule (long above VWAP, short below, flip on every cross) on five years of 5-minute Nasdaq 100 with real spread. It lost 47%, tuning didn't rescue it, and none of the changes we tried made money on six months held back from the tuning.
From r/technicalanalysis. The idea, in our words: The rule from the 2023 VWAP paper the post discusses: in the New York session, go long when price closes above the session VWAP and short when it closes below, flipping on each cross, flat at the close.
no edgeAs posted · the rules as described (see what we assumed below)
Optimized: best settings Quant Core found · last months held back as unseen data
Improved: only trade with the trend (price above/below the 60min 200 EMA)
We then tried 10 changes a quant would test first (800 settings in all) and kept the best on past data only. The held-back months played no part in choosing it. It did not hold up on the unseen months.
Ideas tried for this post
Each idea was tuned on past data only, then run on the held-back months (£ won per £1 lost).
| Idea | Tuning months | Unseen months |
| trade only the first two hours (9:30 to 11:30) | -28% · £0.94 | +1.4% · £1.03 · 217 trades |
| no flipping: take the day's first VWAP cross and hold it to the close | +58% · £1.05 | -44% · £0.28 · 84 trades |
What we had to assume
- Nasdaq 100 CFD with its real spread, instead of QQQ
- 5-minute bars instead of the paper's 1-minute bars
- Long and short (the paper's rule); the post's long-only version is the other reading
- A wide 3 ATR safety stop (the rule has none); flips on the opposite cross; flat at 16:00 New York
A different reading of the rules could give a different result.
Backtested with the StratOS optimizer on historical data, with spread and costs charged, and the most recent months held back as unseen data. A backtest is not a forecast, and nothing here is investment advice.