Why Our Orderflow Momentum Bot Failed

Orderflow looked attractive because it seemed closer to market mechanics than conventional indicators. Our PAPER experiments were a useful reminder that a plausible mechanism is not the same thing as a durable trading edge.

The hypothesis

One branch looked for consolidation followed by a breakout and rapid cumulative-delta acceleration. Another explored order-book imbalance and reversion behavior. The goal was to react to short-horizon buying and selling pressure rather than lagging price indicators.

CONSOLIDATION→ORDERFLOW SHIFT→BREAKOUT→TRADE

What PAPER testing exposed

PROBLEMWHY IT MATTERED

Noisy microstructureShort-lived imbalance did not reliably become a move large enough to support the complete trade.

Execution sensitivityA short-horizon edge has little room for fees, spread, timing error and adverse movement.

False breakoutsAcceleration around a boundary could disappear quickly instead of developing into persistence.

Parameter temptationRepeatedly tightening filters after losses risked fitting the observed sequence rather than improving the hypothesis.

Why we stopped the branch

The correct response to weak evidence was not to keep adding complexity indefinitely. We archived the approach and moved on. A failed branch is valuable when it prevents the same hypothesis from being rediscovered and repackaged later.

What we carried forward

The project shifted toward versioned forward experiments, explicit net-PnL measurement and architectures that make comparison easier. The lesson was not that orderflow can never work; it was that our implementation had not demonstrated an edge worth continuing.

Failure framework →

Continue the project history

From V1 to V6 →Next: V5.1 lessons →