Information report (IC / IR)
python main.py info measures whether a strategy has skill — its information
coefficient and effective breadth — and reconciles the predicted information ratio
with the realized one. It is read-only: a diagnostic, never a control input.
python main.py info \
--strategy volume_spike \
--symbols NVDA,AAPL,META,AMD,TSLA,GOOG,MSFT,AMZN \
--start 2024-01-01 --end 2024-12-31 \
--n-trials 10
Information report: 'volume_spike' 2024-01-01..2024-12-31
measured over 24 rebalances (horizon 5 bars)
IC mean +0.018 t-stat +0.74 rank-IC +0.021
breadth: 142 effective vs 504 naive (ρ̄ 0.41, 8 names)
IR: predicted +0.21 realized +0.18 ± 0.61 (SE)
guardrails: P(any |t|>2 in 10 trials) = 0.40
Verdict: skill is NOT distinguishable from luck (IC t-stat +0.74).
Read it as: the IC t-stat is the honesty gate — below ~2 the mean IC is a few
lucky periods, not skill. Effective breadth deflates the name count by how
correlated the bets are (ρ̄); predicted IR = mean_IC · √BR_eff. The realized
IR comes with a standard-error band — a 1-year window has SE(IR) ≈ 1, so almost
any IR is indistinguishable from zero on a single short window. --n-trials reports
how inflated a "significant" result is once you account for everything you tried.
Options
| Flag | Default | Meaning |
|---|---|---|
--strategy | volume_spike | The strategy whose alpha is measured. |
--source | strategy | Alpha score origin (strategy / signal / scanner). |
--symbols | demo universe | The cross-section. |
--start / --end | last year | Measurement window. |
--benchmark | SPY | Used to strip beta (residual returns). |
--horizon | 5 | Forward-return horizon, in bars. |
--n-trials | 1 | Configs tried, for the multiple-testing inflation. |
--neutralize-factors | off | Measure the factor-neutral alpha (bare flag = market,volatility,size) — use the same setting you deploy with, so the measured IC/IR describes the forecast you actually trade. Also on horizon. |
--scaling-ab | off | Research mode: walk-forward the realized IR under Case-1 (σ·IC·z) vs Case-2 (IC·c_g·z) scaling and compare against the regression's pick — the ground-truth tiebreak for the Case test. |
The level shrink and the risk-bucket monitor
The report also carries two more diagnostics:
- Level shrink. The measured IC is itself estimated; the report prints
what fraction of the naive level survives that estimation error —
keep 13% of the naive level (T_eff 60, IC 0.05)— and the shrunk IC to deploy.T_effdeflates the rebalance count for horizon overlap, so a daily-sampled monthly horizon isn't credited 21× the observations it really has. - Risk buckets. Under correct scaling every residual-vol bucket contributes ~equally to active variance; a monotone gradient flags a mis-scaled alpha (usually a Case mis-choice). Suppressed on universes too thin for reliable buckets rather than reporting noise.
What it does (and doesn't) tell you
- Skill vs luck. A high IC with a low t-stat, or a realized IR inside its SE band of zero, is not skill — the report says so plainly.
- No look-ahead. The IC pairs each forecast with strictly later residual returns.
- Feedback to alphas. When skill is distinguishable, the report recommends the measured IC to replace the prior used in alpha scaling — a human applies it; nothing auto-tunes the trade clock.
The same report is available to agents as the read-only MCP tool compute_information.
Performance attribution (--attribution)
info --attribution replaces the pooled IC/IR report with a per-period,
per-source breakdown of realized active return: systematic benchmark timing,
each risk factor, each signal (your strategy's own alpha, plus any
--attribution-signals you name), and stock-picking — every row exact by
construction (they sum to the realized active return) and Bayesian-blended for
a short-sample-honest t-stat.
python main.py info --attribution \
--strategy volume_spike --symbols NVDA,AAPL,META,AMD,TSLA,GOOG,MSFT,AMZN \
--start 2024-01-01 --end 2024-12-31
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%
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 and the factor rows as cheap
tilts you could have gotten from an ETF. --attribution-signals ma_crossover,mean_reversion attributes other strategies' combined scores as
additional signal columns, so a combined-alpha
weight can be checked against its realized counterpart. Add --conditional ewma|har for a predicted-vs-realized tracking-error table split by volatility
regime, and --bootstrap-skill for a nonparametric own p-value next to the
parametric verdict (see below). See
Performance attribution (engineering) for the
regression identity and why signals and risk factors are fit jointly.
Conditional-risk A/B (--conditional-ab)
The net-of-cost decision for conditional risk: walks the window forward constructing the same alpha book against a conditional and an unconditional Σ, carries each variant's weights forward, and compares realized net IR — not TE-tracking alone.
python main.py info --conditional-ab --conditional ewma \
--strategy volume_spike --symbols NVDA,AAPL,META,AMD,TSLA --start 2024-01-01 --end 2024-12-31
Conditional-risk net-of-cost A/B: 'volume_spike' 2024-01-01..2024-12-31
measured over 15 rebalances (horizon 21 bars, method ewma)
variant net IR realized TE pred TE turnover
unconditional +1.25 12.5% 11.3% 6.7%
conditional +1.23 12.2% 11.2% 6.7%
winner (net IR): unconditional
(net of the real transaction cost — a Σ that tracks TE better but churns the book loses here)
Multi-period-policy A/B (--policy-ab)
The equivalent net-of-cost decision for
multi-period trading: the myopic policy
vs the aim (--policy aim) policy on the same alpha book.
python main.py info --policy-ab --strategy volume_spike \
--symbols NVDA,AAPL,META,AMD,TSLA --start 2024-01-01 --end 2024-12-31
Multi-period policy net-of-cost A/B: 'volume_spike' 2024-01-01..2024-12-31
measured over 12 rebalances (horizon 5 bars)
variant net IR realized TE pred TE turnover
myopic -1.07 1.5% 1.2% 8.3%
aim -1.07 1.5% 1.2% 8.3%
winner (net IR): myopic
--trade-rate overrides the derived κ for the aim leg. over_damped (shown
as a warning line when it fires) means the aim policy traded less and still
scored a lower net IR — not an improvement, over-damping.
Nonparametric skill check (--bootstrap-skill)
Adds a stationary block-bootstrap own p-value (no distributional assumption on
returns) next to the parametric SE{IR}≈1/√Y verdict — the heavier,
definitive check behind the report's skill-vs-luck call. See
Evaluation metrics — bootstrap skill inference
for the family (Reality Check) half of this, which needs the
trial store and lives on
walkforward --bootstrap-skill instead.
Not yet covered
Capacity analysis lives on the portfolio construction report
(capacity_capital), not here.