Six years of open interest were sitting in a public archive
- Published
- August 26, 2026
- Reading time
- 2 min
- Topics
- open interest · data archaeology · statistical power
The investigation’s bottleneck was supposed to be the calendar. With the price-and-volume family closed, the next candidates are microstructure signals — open interest, positioning ratios, book depth — and a test on them only means something after roughly a year of history. The collectors were months in. Then it turned out most of that history already exists, in public, for free.
01What was found, and what was checked
The exchange’s public archive [1][2] publishes daily files that include open interest and long/short positioning ratios at five-minute resolution, going back to September 2020 for the oldest pairs. Samples were downloaded and the schema inspected directly: the fields match what the live collectors capture, at twelve times the sampling density of the system’s own hourly polling. For the two signals closest to maturity, that is roughly six years of test power instead of nine more months of waiting.
The same archive holds order-book depth from January 2023 — usable, but at percentage bands rather than the level-by-level shape the live collector records, so any hypothesis on it has to be defined on features both sources can compute. The honest misses: the best-bid/ask feed needed to reconstruct true order-flow imbalance was discontinued in March 2024, and there is no free historical source for liquidations. Those two signals keep accumulating in real time.
02What this does and does not change
Before any hypothesis touches this data, the archive has to earn trust: the window where archive and live capture overlap gets audited for divergence — timestamps, units, sampling artifacts. Only after that does the open-interest family get its pre-registered hypotheses, under the same gates as the fifty that died: permutation p<0.01, out-of-sample, deflated Sharpe, beat equal-weight net of costs.
And to be clear about what this is not: it is data, not edge. The count of validated strategies remains zero. What changed is that the next round of experiments starts years earlier than planned — which cuts both ways, since the most likely outcome of more statistical power is, as always, a faster funeral.
Numbered sources the body points at. The order is fixed at publication — it is part of the text.
- GitHub — binance/binance-public-data
- Binance