Why the Kalman pairs research failed
Two frozen versions of an altcoin pairs strategy hedged altcoin futures against each other with Bitcoin exposure removed, then attempted to trade the residual spread on reversion. Both versions were paper-tested. Both lost after costs. The mechanism is straightforward, the measurements are clean, and the finding is honest: a tighter hedge is not an edge.
The Premise and the Mechanism
The strategy takes two altcoin futures and removes their shared Bitcoin exposure by hedging each against Bitcoin, leaving a spread between the two altcoins. The expectation is that this spread reverts after temporary dislocations.
The hedge ratio between each altcoin future and Bitcoin was not fixed. A Kalman filter re-estimated it on every 15-minute bar, meaning the hedge adapted continuously as the relationship between each altcoin and Bitcoin shifted. The tighter the filter's tracking, the more precisely Bitcoin exposure was removed and the smaller the residual spread became. That precision is what the first version's failure turned on.
The First Version: Zero Trades
The initial configuration set the Kalman filter's process-noise delta at a level that let the hedge ratio move quickly. Over 168,000 model fits across the research window, the strategy generated 0 trades.
The cause was mechanical, not philosophical. The hedged spread never left its entry band. The filter tracked the altcoin-to-Bitcoin relationship so closely that Bitcoin's influence was stripped out aggressively, leaving a residual spread that stayed inside the threshold required to open a position.
The dispersion measurement confirmed this directly. The median deviation of the spread from its entry threshold was 0.129 of the level needed to enter. In other words, the typical bar came nowhere close to triggering a trade, because the hedge was absorbing nearly all the movement that would have created a signal.
The cleanest possible finding from this version is that a hedge tight enough to eliminate Bitcoin's influence also eliminates the spread's ability to signal a dislocation. The strategy had no mechanism to trade.
The Single Change
One parameter changed: the process-noise delta was reduced from its original value by a factor that moved it to a slower-tracking regime. The filter now updated the hedge ratio more conservatively, allowing the spread to move more freely. Nothing else in the design changed.
The effect on dispersion was immediate and large. The median deviation rose from 0.129 to 0.474 of the entry threshold. The spread now moved far enough to reach the band, and the strategy could finally open positions. Whether those positions would make money after costs was a separate question, and the research was designed to answer it.
The Follow-Up: Trading and Losing
The follow-up was tested on 13 folds from 2025 to 2026, across 14 altcoin futures against Bitcoin, with a static-OLS control on the same data. Both versions were evaluated on unseen folds.
The Kalman version produced 118 unseen trades. The win rate was 33.1%. Net per trade after the 16 bps round-trip cost was −50.8 bps. The profit factor was 0.32.
The static-OLS control, which used a fixed hedge ratio rather than a Kalman filter, produced 125 trades at a 30.4% win rate, −55.8 bps net per trade, and a profit factor of 0.33.
In-sample variants ranged from −43.9 to −57.3 bps per trade, with a profit factor at most 0.39. Only 1 of the 14 pairs was profitable on its own.
The loss was not driven by a single event. The largest single Bitcoin impulse carried 4.96% of the absolute loss, meaning the damage was distributed across many events rather than concentrated in one accident.
The Control and the Conclusion
The Kalman hedge improved on the static-OLS control by approximately 5 bps per trade. Both versions lost money after costs. The improvement in hedge quality did not translate into an edge.
The premise is dead at this cost level on this universe and window. A tighter hedge is not an edge. The strategy failed, and the failure is attributable to the mechanism itself rather than to one parameter or one period.
The holdout window from the 2025 to 2026 test design was deliberately left unspent. It remains available for a successor strategy to test once, preserving the integrity of the research record.