Zach Anderson
Mar 13, 2026 17:07
New analysis reveals backtests utilizing revised on-chain knowledge produce deceptive outcomes. Level-in-time metrics reveal considerably worse real-world efficiency.
That worthwhile buying and selling technique you backtested? It most likely would not have labored in actual time. Glassnode’s newest analysis demonstrates how retroactively revised on-chain knowledge creates a harmful phantasm of profitability that evaporates when examined in opposition to info merchants really had entry to.
The analytics agency ran similar backtests on a easy BTC alternate stability technique—one utilizing normal historic knowledge, one other utilizing immutable point-in-time (PiT) metrics. Similar sign logic, identical parameters, identical 0.1% buying and selling charges. The outcomes diverged dramatically.
The Hidden Downside with On-Chain Knowledge
Metrics like alternate balances aren’t static. They get revised as deal with clustering improves and entity labeling updates. That Binance BTC stability determine you are taking a look at for January 15, 2024 could not match what was really printed on that date.
Once you backtest in opposition to revised knowledge, you are buying and selling on info that did not exist when selections would have been made. This look-ahead bias is especially extreme for metrics depending on entity identification—precisely the form of knowledge many merchants depend on for alternate stream evaluation.
Glassnode’s check technique was simple: go lengthy when the 5-day transferring common of Binance’s BTC stability drops under the 14-day common (sustained outflows), exit when it crosses again above (outflows reversing). Operating from January 2024 by means of March 2026 with $1,000 preliminary capital, the usual backtest confirmed efficiency roughly matching buy-and-hold.
The PiT model informed a unique story. Whereas each methods tracked equally by means of a lot of 2024, the immutable knowledge model missed the sturdy November 2024 and March 2025 rallies that the revised-data backtest captured. Cumulative efficiency ended up “significantly decrease,” in response to Glassnode.
Why This Issues for Quant Merchants
The implications prolong past this single technique. Any backtest counting on knowledge topic to revision—alternate balances, entity-tagged flows, even buying and selling volumes from exchanges that report with delays—faces the identical contamination danger.
This aligns with broader considerations in quantitative finance about knowledge high quality. Analysis from different knowledge suppliers reveals PiT methodology prevents a number of bias sorts: look-ahead bias from utilizing future revisions, survivorship bias from datasets that exclude failed entities, and hindsight bias from restated figures.
For crypto particularly, the place on-chain analytics corporations constantly refine their entity labeling and clustering algorithms, the revision drawback compounds. A pockets recognized as belonging to Binance as we speak won’t have been tagged accurately two years in the past when your backtest assumes you traded on that sign.
The Sensible Repair
Glassnode now gives PiT variants for all metrics by means of their Skilled tier. These append-only datasets lock in every knowledge level because it was initially computed—no retroactive adjustments.
The tradeoff is actual: your backtests will doubtless look worse. However they’re going to mirror what would have really occurred. For merchants allocating actual capital primarily based on quantitative alerts, that accuracy hole between a flattering backtest and disappointing dwell efficiency could be costly.
Earlier than deploying any technique constructed on on-chain metrics, the query is not whether or not the backtest seems to be worthwhile—it is whether or not you examined in opposition to the info you’ll have really seen.
Picture supply: Shutterstock
