flow-intel desk
A self-learning desk that tests which flow signals actually validate out-of-sample.
Agent token 9. flow-intel is a self-learning desk: it doesn't assume any flow signal works — it continuously tests a basket of them (OI, funding, order-flow, momentum) and only lets the survivors earn weight. It's the desk that embodies the firm's "an edge must be earned" rule.
What it reads
firm/flow_intel.py — a rotating set of flow-signal hypotheses, each forward-logged and scored:
- OI — open-interest change behaviour.
- Funding — perp funding-rate extremes.
- Order-flow — aggressor imbalance.
- Momentum — short-horizon continuation.
Each candidate is held to a cost-aware, out-of-sample standard and a Benjamini-Hochberg FDR gate
so that running many tests doesn't manufacture a false winner. On the data in this repository 0 of 4
validate for NEAR (and for BTC): the stored positioning history is too short for the test, so every signal
reads insufficient data and gets zero weight. The desk reports that honestly rather than forcing a signal.
Signal lifecycle
Why FDR matters
If you test 20 signals at p < 0.05, you expect ~1 to look "significant" by pure chance. A naive
best-of-N search guarantees a false edge. The Benjamini-Hochberg procedure controls the
false-discovery rate across the whole battery, so a signal that clears the gate is far likelier to be
real. This same FDR discipline is shared with edge_lab and the onboarding screen.
Honest caveats
flow-intel is mostly a producer of "no" — and that's the point. It exists to stop the firm from
trading flow folklore. When it does validate a signal, that signal folds into desk_performance and is
ranked alongside the other desks; until then it abstains. See Self-learning.