← Research logMomentum & Time-Series
KILLED · latency decay
#

The Cont–Kukanov order-flow imbalance at the best bid/ask linearly predicts the next mid-price move; a fast book-pressure signal captures sub-minute drift.

OFI aggregates signed size changes at the top of book over $[t-\Delta,t]$:

$$ \text{OFI}_t = \sum_{k} \big(\Delta q^{b}_k\,\mathbb{1}_{\Delta P^b\ge 0} - \Delta q^{a}_k\,\mathbb{1}_{\Delta P^a\le 0}\big) $$

Predict the mid-move linearly:

$$ \Delta \text{mid}_{t,t+\delta} = \lambda\,\text{OFI}_t + \eta_t $$

Reconstruct L1 book from the public depth stream, regress next-$\delta$ mid-move on OFI for $\delta\in\{1,5,15\}$s. Net-of-fee replay assuming taker entry.

In-sample $R^2$ ($\delta$=1s)0.21
Predicted move per signal0.6–1.4 bps
Round-trip taker cost~7 bps
Net EV after fees + latency
KILLED
The signal is genuine ($R^2\approx0.2$) but the predicted move (~1 bp) is an order of magnitude below the fee. Capturing it requires maker fills and colocation; at public-API latency and taker fees it is dead. A market-maker’s edge, not a taker’s.
Predictive power and tradeable edge are different quantities. A signal worth 1 bp is worthless to anyone paying 7 bps to act on it.

We publish the failures too.

This is one of 100+ documented hypotheses. Browse the full lab notebook, or see the strategies that survived into production.