Machine-learning strategies and event data

Two leaks that no line of code shows: a model fitted on the whole file, and events joined by the time they happened instead of the time they became public. Monte-Neo takes both decisions away from the strategy.

The fit / predict protocol

A model fitted inside signal(df) on every row still reacts to its input row by row, so the look-ahead probes cannot see that it learned from the future. Define two functions instead of signal:

import numpy as np

REFIT_EVERY = 63     # bars between refits
MIN_TRAIN = 252      # bars before the first prediction
LABEL_HORIZON = 1    # bars a training label looks ahead
EMBARGO = 5          # extra bars between the training rows and the first predicted row


def fit(train):
    """Return any object: a fitted estimator, coefficients, a dict."""
    ...


def predict(model, df):
    """One position per row of df (+1 / 0 / -1 or weights)."""
    ...

monte-neo verify --ohlcv data.csv --strategy model.py

The verifier runs the walk-forward itself. Each block of REFIT_EVERY bars is predicted by a model fitted on rows [0, start − LABEL_HORIZON − EMBARGO), so no prediction comes from a model that saw its own future. The look-ahead probes then test predict. Fitted models are cached by their exact training rows, so the probes do not refit for nothing.

Check	What it says
ml_refit_gap	The Sharpe with honest refitting against the Sharpe of predict(fit(df), df), the usual whole-file backtest. fail (look-ahead) when the in-sample edge is real and the walk-forward loses money: the profit lived in the fit. warn when less than half the edge survives.
ml_cv_overlap	If the file defines cv_splits(df) (a list of (train_rows, test_rows)): training rows whose labels reach into a test fold, or that sit inside the embargo after it. Purge them.

Two more checks are optional:

• ml_seed_stability: set SEEDS = 5 and give fit a seed argument. The walk-forward runs once per seed, and a result whose sign or size depends on the seed is flagged.
• ml_feature_drift: define features(df). The population stability index of each feature between the first training window and the predicted period is reported, with a warning above 0.25.

Two lint rules flag the classic leaks of machine-learning code written inside signal(df): full_sample_transform (a scaler, PCA or feature selector fitted on all rows before any split) and target_encoding_full (a mean of the target over all rows mapped back to the features). Both are warnings: the protocol above is the real fix.

The constants are part of the strategy source, so they are hashed into the certificate. Setting LABEL_HORIZON or EMBARGO lower than the truth weakens the test, and anyone who reads the source can see it.

Events joined by publication time

Earnings surprises, news sentiment, alternative data and LLM outputs reach a strategy as events. The common leak is a join on the event's own time (the quarter end, the moment the fact happened) instead of the time it was published.

monte-neo verify --ohlcv data.csv --strategy s.py --events events.csv [--events-delay 3600]

events.csv has published_at (required), event_time (optional), symbol (for universes) and numeric value columns. The verifier joins them to the bars as ev_<column> columns, and the strategy reads those.

• An event is visible from the first bar that opens after published_at + delay. A bar that was already open when the news came out could not have used it.
• Nothing is filled before the first event.

Check	What it says
events_quality	Events published before they happened, a zero publication delay (is published_at really the publication time?), events outside the price range
pit_events	The same strategy run once more on a join by event_time. A warning when that join earns more than twice as much and 5 points more: most of the edge would need knowing events before they were public.

The events go into the certificate as a hash and a delay. Recheck with the same file: monte-neo verify --recheck cert.json --ohlcv data.csv --strategy s.py --events events.csv.

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