Part two of the breakout series.
Last time we said we were training a model to fill in trade-by-trade detail from one-minute bars, with grades fixed
before the first run. Here are the grades - all fifteen fields, including the ones that failed.
How it was graded
- Trained on 150 days; graded on 90 days from nine different market moods it never saw (a calm bull market, the 2018 selloff, the 2020 crash, the 2021 retail mania, the 2022 bear market and four more), with a month of distance between training and test days.
- The score is R2: the share of the real minute-to-minute variation the model explains on those unseen days.
- What it can't cover: a minute with trades but no one-minute bar can't be filled from a bar. Depending on the day that is 8-30% of traded minutes but only about 0.1-0.3% of traded volume - tiny minutes, typically 15-26 shares. The grades cover the minutes that carry about 99.8% of the volume.
- GOOD = at least 0.60, PARTIAL = 0.25 to 0.60, POOR = under 0.25. Fixed before training.
The grades
GOOD - 7 of 15 fields, and they held up in every one of the nine moods:
| Field | R2 |
|---|---|
| Largest single trade's share of the minute | 0.88 |
| Bid-ask spread | 0.86 |
| Odd-lot share | 0.80 |
| Retail share | 0.74 |
| Quote depth | 0.69 |
| Block-trade share | 0.63 |
| Quote count | 0.61 |
PARTIAL - 5:
| Field | R2 |
|---|---|
| Locked / crossed quotes | 0.51 |
| Order-flow imbalance | 0.46 |
| Buy share | 0.39 |
| Off-exchange share | 0.36 |
| Retail buy-minus-sell | 0.28 |
Buy share is worth a second look: the classic shortcut that estimates buying from the bar alone (bulk volume
classification) explains 0.09. The model gets more than four times that - and still only a partial grade.
POOR - 3: bid-minus-ask depth (0.03), first-ten-seconds share of the minute (0.04), intermarket sweep share (0.18).
The catch that matters
The most important number isn't on the table. At the 737 real breakout minutes in the test days - the exact
moments the exit question is about - the model explains only 6% of buy share, and it underestimates it: real buy
share at those minutes averaged 71%, the model said 63%.
It isn't only buy share. Retail share and off-exchange share are fine on a typical minute but weak at the breakout
minutes (R2 0.21 and 0.20). The structural fields hold up there: spread, odd-lot share, block share and largest-trade
share stay between 0.77 and 0.87.
That makes sense once you see it. The buying pressure that made those minutes special is exactly what a one-minute
bar cannot show. A model built from bars learns the ordinary minute well; it smooths over the unusual one.
What we do with this
- The GOOD fields are usable for what we said: realistic spreads, depth and trade-size structure in simulations and training environments where real ticks don't exist.
- Pressure-based exits get tested on real ticks only. The generator is not good enough at the minutes that matter to stand in for them.
- And the general rule stands: generated fields add no information a bar didn't have. Anything learned on them is confirmed on real data before it means anything.
One process note, because it is the kind of thing that quietly ruins results: the first training sample had a bug - for
2024-2025 it read one file per day where some days have many. It was caught before any result was used, the sample was
rebuilt, and everything was regraded. The grades above are from the rebuilt sample.
A question for you
If you have estimated buy pressure from bars, how did you score it - and did it hold up at the minutes that mattered
most to your strategy?
Research and education only - not investment advice.
Top comments (0)