Skip to main content

Performance attribution

Information analysis checks the predicted IR against the realized one, pooled over a window. tradeflow/analytics/attribution.py (+ services.compute_attribution) goes further: it attributes realized active return per period, per source — systematic benchmark timing, each risk factor, each signal, and stock-picking — by an exact regression identity, then puts the same research-integrity guardrails on the attributed series that information analysis puts on ICs.

A backtest tells you what happened. Attribution tells you where it came from — timing the benchmark, riding a factor, or genuine name selection — which is what tells you whether it repeats.

Research clock only

A read-only diagnostic over realized returns; it places no orders and feeds nothing back automatically.

The regression identity​

At each sampled rebalance, active return splits into two stages, each an exact regression identity so nothing is double-counted or lost:

  1. Systematic (known, predetermined) benchmark timing. Strip βᵢ(t)·r_B(t) per name using the same canonical Σ-implied beta risk model already computes, so the portfolio aggregate is exactly β_a(t)·r_B(t) — not fit to this period's data, just the already-known tilt times the realized benchmark move. Decomposed further into expected / surprise / timing.
  2. Risk factors + signals, jointly. One cross-sectional regression of the beta-adjusted return on [risk_factor_exposures, signal_scores] together.

The two blocks are fit jointly, not signals as a second pass on the risk-factor residual. A second pass looks cleaner (each block seems to "own" its regression) but leaves an omitted-variable bias that does not vanish with more names: when active weights are built from a signal that isn't yet among the regressors doing the fitting — exactly how a paper book is built upstream, from its own alpha z-score — whichever risk factor happens to correlate with that signal in this period's finite sample picks up a persistent, same-sign share of the signal's true return. That's a genuine integrity bug (measured well past significance in development, not sampling noise around zero), not a rounding error; fitting both blocks together removes the omission and the bias disappears. The real remaining tradeoff — signals correlated with risk factors destabilize a joint fit — is smaller and disclosable (the condition number is reported); the omitted-variable bias was not disclosable, since it looks exactly like real factor timing.

Every component sums back to r_active exactly (up to floating point) — residual in the output is adding-up slack only, and a test pins it near zero.

Honest, per-row statistics​

Every attributed row (systematic, each factor, each signal, specific) gets its own t-stat, using a Bayesian-blended risk: short attributed series lean on the risk model's structural prior instead of a wild few-point sample standard deviation, blending toward the realized variance as the sample grows relative to a prior weight T₀ (derived from the risk model's own min_obs, converted to this attribution's rebalance-period units — "the risk model itself wouldn't trust an estimate from fewer periods than this, so neither should the attribution").

A ranked table of ~8 attributed rows is exactly the same multiple-testing trap a ranked IC table is — quoting the single best row without correcting for how many you looked at overstates significance. The whole table is deflated by the same P(any |t| > 2 in n_trials) inflation information analysis applies to ICs.

Honest (non-compounding) cumulation​

Chain-linking a per-period additive split across many periods is not the same as compounding the true portfolio and benchmark returns — treating it as if it were is the cumulation trap. Each component is chain-linked by the compounding path up to (not including) its own period, which telescopes so the linked components sum to Π(1+r_active) − 1 exactly (naive_cumulative). The gap between that and the honest ΠR_P − ΠR_B (honest_car) is delta_cp, reported — not hidden — and flagged cumulation_unreliable when it's large relative to the attributed terms (large per-period returns compound nonlinearly enough that the additive story stops being a good local approximation).

Optional: conditional risk, extra signals, bootstrap skill​

  • conditional threads an EWMA/HAR-conditioned Σ(t) into the per-period covariance this function already rebuilds at every sampled rebalance; when set, the report adds te_by_regime — predicted TE vs a realized-dispersion proxy, bucketed by the benchmark's own trailing realized-vol tercile. See Risk model — conditional risk. One Σ choice per call; the net-of-cost conditional-vs-unconditional comparison lives in info --conditional-ab.
  • signals attributes extra strategies' combined scores as additional signal columns (the strategy's own alpha is always included) — so a multi-signal combination weight can be checked against its realized counterpart.
  • bootstrap_skill adds a nonparametric OWN p-value next to the parametric SE{IR}≈1/√Y verdict — see Evaluation metrics.

Example​

Performance attribution: 'volume_spike' 2024-01-01..2024-12-31
  measured over 24 rebalances (horizon 5 bars)
  row                       mean/yr      IR       t  share ψ²
  active beta · expected          —       —       —  (not skill)
  active beta · surprise          —       —       —  (not skill)
  timing (δβ·δr)               0.12    0.08    0.31       0.4%
  market factor                -0.31   -0.22   -0.84       2.1%
  momentum factor               0.44    0.31    1.20       3.8%
  specific (stock-picking)      1.02    0.71    2.76      41.2%

  cumulative active return: +2.14% (top-down parts +2.09% + δ_CP +0.05%)
  Verdict: IR +0.68 ± 0.31 — distinguishable from luck (t +2.19)

Read "specific" as genuine stock-picking skill, factor rows as cheap tilts you could have gotten from an ETF, and the systematic rows as the part explained by already-known beta times the benchmark's realized move — not a bet, an accounting fact.

Where it runs​

services.compute_attribution samples rebalances, rebuilds a leakage-safe cross-section at each (alpha, risk-factor exposures, per-period Σ), and attributes. There is no persisted weights/exposure history yet, so — like compute_information — it recomputes per-period from stored bars rather than consuming one. The CLI is python main.py info --attribution; see the usage guide.