Summary
MultiPeerOrderFlowImbalance extends cross-asset OFI impact to a basket of peers.
Each tick, peer OFI values are averaged with equal weights into peer_mean; rolling OLS
of own return on that mean yields beta, impact, and residual.
Update API
import numpy as np
import rtta
ind = rtta.MultiPeerOrderFlowImbalance(window=50, fillna=True)
result = ind.update(own_return, peer_ofis) # peer_ofis shape (n_peers,)
# result.beta, result.impact, result.residual, result.peer_mean
# Batch: own_return (T,), peer_ofis (T, n_peers)
batch = ind.batch(own_returns, peer_ofi_matrix)
advance(...) updates state without returning a result. Empty peer vectors yield NaNs.
Theory Of Operation
Multi-name flow often arrives as a vector of peer OFIs. An equal-weight reduction
acts as a single cross-impact regressor. Rolling \(\beta\) of own return on \(\bar{f}\)
matches CrossAssetOrderFlowImbalance with peer feature \(\bar{f}_t\). This is a lightweight
basket alternative to weighted multi-peer variants.
Recurrence
For peer vector \(f_t \in \mathbb{R}^P\) (\(P \ge 1\)):
Push \((x_t, y_t) = (r_t, \bar{f}_t)\) into RollingPairStats of length \(W\). With
rolling sums \(S_x,S_y,S_{xy},S_{y^2}\) and \(n = |W_t|\):
If fillna=False and the window is incomplete, \(\beta\), impact, and residual are NaN;
peer_mean is still \(\bar{f}_t\).
Implementation Notes
The recurrence is implemented in src/rtta/indicator.cpp in
class MultiPeerOrderFlowImbalance. Batch requires at least one peer column. Float32 and
float64 peer vectors are supported.