Alpha validation: does a ranking signal predict returns, and can it be traded?

monte-neo alpha (MCP tool validate_alpha) checks a cross-sectional score: a number per symbol and date that should rank the returns that follow. It answers two questions a fund's research committee asks before a signal gets capital:

1. Does the score predict anything? (family A, signal power)
2. Can its portfolio be traded at the size you plan? (family D, implementability)

monte-neo alpha --demo                                              # no files: a weak real edge and a noise score
monte-neo alpha --ohlcv universe.csv --scores scores.csv --aum 5e7  # your data and the capital you plan to run

Inputs:

• --ohlcv: one long table for the universe with timestamp, symbol, close and, for family D, volume.
• --scores: timestamp (or date), symbol, score. The score at t ranks the returns after t. Scores are never filled forward, and a score dated after the last price is an error.

The output is an alpha-report/1 report with a verdict, the checks, metrics and next actions.

Family A: signal power

Check	What it measures	Status
ic	Mean rank (Spearman) IC per date against the next return, ICIR and a Newey-West t-statistic	info when the mean IC is positive and t ≥ 2, else warn
ic_decay	IC at horizons 1 to 20 bars, the best horizon and the half-life	info
quantile_monotonicity	Do returns rise from the worst to the best quantile, or does all the profit sit in one tail?	pass or warn
long_short_legs	Share of the spread earned by the short leg	warn above 70 %
breadth_consistency	A claimed IR (--claim-ir) against IC × √breadth	warn above twice the implied IR
signal_placebo	200 scores permuted among symbols within each date: does the real score beat them?	fail when p > 0.10
selection_vs_timing	Is the spread stock selection, or a fixed tilt that survives shuffled dates?	info

Family D: implementability

The portfolio is long the top quantile and short the bottom one, with equal weights.

Check	What it measures	Status
market_impact	Square-root impact at --aum (trailing 20-bar volatility and dollar volume, no look-ahead)	fail when the net return is ≤ 0, warn when impact eats more than half the gross
participation_limit	Your trades as a share of each bar's volume at --aum	warn when the 95th percentile is above 5 %, fail above 20 %
capacity_curve	Net Sharpe from 100 thousand to 10 billion: where it halves and where it reaches zero	info
short_feasibility	Short exposure in the least liquid names, and borrow (300 bps a year unless stated) against the short leg's profit	warn
turnover_vs_decay	Holding period against the IC half-life	info
rebalance_timing_luck	The same portfolio rebalanced every 21 bars on each of the 21 possible days	warn when the spread of results exceeds the median

Without --aum, market impact and participation are skipped and the report says why.

Families B, C, E, F, G and I: on the score's long-short portfolio

These families judge the returns of the same top-minus-bottom portfolio, before costs: they ask whether the signal is real. Costs and size are family D's job.

Family	Checks	Needs
B, attribution	factor_exposure, residual_alpha (is it alpha after market, size, momentum, reversal and low-volatility factors? t ≥ 3 passes), factor_spanning, hidden_short_volatility (steady gains and rare crashes), market_timing_beta, crowding, known_anomaly_match	factors are built from your universe; --factors file.csv uses yours
C, value to your book	book_correlation (also in the book's worst bars), marginal_sharpe (the book's Sharpe gain, with a bootstrap interval), residual_vs_book, drawdown_overlap; signal_orthogonality (partial IC after the signals you already run)	--book book.csv (date + one column per strategy), --existing-scores scores.csv; skipped without them
F, all the attempts	hlz_hurdle (t ≥ 3), min_backtest_length, drawdown_plausibility, haircut_sharpe	--n-trials
I, risk	neutrality, concentration, leverage_gross_net, historical_stress_replay (named crises found in your dates), tail_risk	—

| E, robustness | universe_subsampling (70 % subsets, top-5 share), group_split (--groups), start_date_sensitivity, cpcv_paths, synthetic_market_null (bootstrapped markets without the score's link), noise_injection, data_source_replication (API: a second price source), bar_construction | — | | G, data and process | universe_lookahead (--membership; without it, a warning when the universe looks like today's members), classification_lookahead, cross_market_close_alignment (--exchanges), post_discovery_decay (--discovery-date), research_hygiene | optional files |

--signal-only runs families A and D only. Every report ends with an Alpha Readiness Level, for example ARL 6/9 — blocked at 7: tradable and robust (rebalance_timing_luck).

Reading the demo

The demo's real score has a rank IC of about 0.05 with t ≈ 9 and beats the placebos. It still gets REJECT with --aum 5e7, because daily rebalancing of illiquid names at that size costs more than the edge. The noise score fails the placebo test. That is the point of the command: a significant IC is not yet a tradable alpha.

Honest scope

• Self-built factors are proxies: size is ranked by traded value, and there is no value factor without fundamentals. Pass --factors with academic or vendor factors for a stricter attribution.
• Impact uses the square-root model with a coefficient of 1: a model, not a measurement.
• The score is taken as given. To check the code that makes it for look-ahead, verify that code with monte-neo verify.

Not investment advice: a research example, not a call to trade or a promise of profit. Simulated results do not predict future returns.