Historical study
Does book imbalance predict which way the price moves?
In this historical sample: 55.5% directional hit rate at five minutes, with a descriptive Wilson lower bound of 55.2%.
Out of 205,835 firings the book pointed the right way 55.5 percent of the time, against the 50 percent a coin flip would give. Even the cautious end of the interval sits at 55.2 percent under the interval's independence assumption. Repeated and overlapping observations limit that assumption. In 35.4 percent of firings the price actually moved at all; the rest sat still.
Data · 205,835 observations · 370,423 book snapshots · 4,489 tokens · 11 days · 2026-07-18 to 2026-07-28
| Hit rateshare of firings where the book moved the predicted way | 55.5 % |
|---|---|
| Wilson lower bound (95%)descriptive bound assuming independent firings; clustered uncertainty remains untested | 55.2 % |
| Firings | 205,835 observations |
| Of those, price actually moved | 35.4 % |
Next test · planned
Does a frozen imbalance rule improve prediction on new events after latency, with uncertainty clustered by day and event?
Research renewal planPayload snapshot 2026-10-12. These figures are frozen; the interactive page reads the same file.
Page export dates do not extend the observation period shown above.
Full report (docs/research/orderflow_rest-2026-07.md) · Source module (src/orderflow_study.py)
Research tool only: no investment advice, no order placement, no venue affiliation. Public Polymarket and Kalshi data, provided as-is.