Alpha Capital

A research factory, not one strategy

Each research direction starts from a literature-backed market mechanism, then runs through Alpha Capital's own point-in-time, survivorship-corrected, cost-aware pipeline. The papers motivate hypotheses; they do not substitute for validation. Most are supposed to die in that pipeline; the value is a system honest enough to tell which few don't.

Literature-backed, not literature-proven: the citations define mechanisms and known failure modes; Alpha Capital re-tests every hypothesis against its own leakage-guarded, cost-aware validation stack.

Six providers, one point-in-time evidence spine

Alpha Capital normalizes market data, listings, filings, news, and quote evidence into a lineage-tracked PostgreSQL store. Each row is tied back to the provider attempt that produced it, so the system can distinguish a real no signal from missing data, a provider failure, a parser failure, or an intentional exclusion.

Scale

Parameterizable U.S. equity universe; the Capitulation-Volume Bounce worked example focused on $30M–$5B small caps.

Five accepted corpora

Data-integrity states

These states are kept separate so bad data cannot silently become a clean negative example.

State	Meaning
Valid observation	Fetched, parsed, and normalized successfully.
Provider fetch error	The API call failed or returned an error; the attempt is recorded, not a silent absence.
Parser / normalization error	Payload arrived but could not be cleanly mapped to the schema.
Missing-but-expected	A source attempt was expected but produced no row.
Intentional exclusion	Filtered out by rule (non-common, ETF, ineligible security type).
No signal / did not qualify	Clean data that simply did not meet the pattern's conditions.

Why this is hard

Most retail backtests use today's ticker list and clean historical bars. Alpha Capital instead reconstructs what the market looked like on each historical date, including delisted names, provider gaps, split adjustments, stale quotes, and failed fetches. That is the difference between a chart and an auditable dataset.

The evidence spine, in SQL

The point-in-time and fail-closed guarantees are enforced in the queries and the schema, not just described. Two representative examples from the repository:

i12_pit_ is the internal table prefix for the point-in-time worked-example corpus (the Capitulation-Volume Bounce); the scratch-schema qualifier is omitted below for readability.

WITH scoped AS (
  SELECT i12_pit_candidate_id
  FROM i12_pit_candidates
  WHERE is_active IS TRUE
    AND candidate_status = 'passed'
    AND decision_date BETWEEN :start_date AND :end_date
    AND decision_time_label = :decision_time
    AND path_mode = :path_mode
), counts AS (
  SELECT scoped.i12_pit_candidate_id,
         COUNT(child.i12_pit_quote_replay_id) AS row_count
  FROM scoped
  LEFT JOIN i12_pit_quote_replays child
    ON child.i12_pit_candidate_id = scoped.i12_pit_candidate_id
   AND child.quote_role = :role
   AND child.is_active IS TRUE
  GROUP BY scoped.i12_pit_candidate_id
)
SELECT
  SUM(CASE WHEN row_count = 0 THEN 1 ELSE 0 END) AS missing_count,
  SUM(CASE WHEN row_count > 1 THEN 1 ELSE 0 END) AS duplicate_count
FROM counts;

Scopes to active, passed candidates inside the decision window, then verifies each required quote role has exactly one active evidence row. Any missing or duplicate count halts the tape build instead of silently producing an ML artifact.

CREATE TABLE historical_universe_reconstructions (
  replay_date        DATE    NOT NULL,
  normalized_symbol  TEXT    NOT NULL,
  inclusion_status   TEXT    NOT NULL,
  delisted_date      DATE,
  reconstructed      BOOLEAN NOT NULL DEFAULT true,
  input_hash         TEXT    NOT NULL,
  output_hash        TEXT    NOT NULL,
  -- ... lineage, provenance, and job-run columns ...
  CONSTRAINT ux_historical_universe_recon_date_symbol
    UNIQUE (replay_date, normalized_symbol)
);

One reconstructed row per (historical date, symbol) — the structural guarantee behind survivorship-correct point-in-time universes, delisted names included. Condensed from the Alembic migration; the production table also carries provenance, lineage, job-run, foreign-key, and index fields.

The public page and GitHub repository show derived / aggregate artifacts only; raw licensed market data and credentials are excluded.

The discipline is the product

Every control below exists to make sure the data cannot quietly lie.

The ML layer: score candidates, then let validation decide

Alpha Capital uses scikit-learn models as selection engines, not as proof of an edge. A model scores each candidate using only point-in-time features; the book is then judged out-of-sample after spread, slippage, tradeability, and cash handling.

Model card

OOS readout

Research-shadow prototypes: the signal was real enough to pursue

Research-shadow onlyNot promoted

Before the final point-in-time, cost-aware gate, I ran several research-shadow ML tests to check whether the model could rank candidate events better than hand rules. These were not production claims: some were PIT-deferred or prototype runs, and later validation was stricter. But they showed meaningful rank separation and justified building the full leakage-controlled pipeline.

These were research-shadow evidence, not promotion evidence. The later point-in-time, survivorship, spread, slippage, and walk-forward gates reduced or blocked the strategy. That is the point of the platform: promising models are not allowed to graduate unless the full evidence chain survives.

Oracle diagnostic: predicting the forbidden full-day-volume signal

Predictors vs. labels: the firewall

The rule: future and full-day data can define an outcome to predict, but can never enter the feature vector. Any predictor that crosses the line is rejected before training (the leakage audit below enforces it).

Implementation proof

These snippets aren't decorative; they're the enforcement points in the repository's real scikit-learn pipeline, not an AI simulation or notebook toy:

I12 is the codebase's internal name for the Capitulation-Volume Bounce, so it appears in the constant names and file paths below.

def _new_gbrt_model(*, max_iter, random_state, model_params=None):
    from sklearn.ensemble import HistGradientBoostingRegressor

    params = dict(model_params or {})
    return HistGradientBoostingRegressor(
        max_iter=max_iter,
        min_samples_leaf=int(params.get("min_samples_leaf", DEFAULT_MIN_SAMPLES_LEAF)),
        l2_regularization=float(params.get("l2_regularization", DEFAULT_L2_REGULARIZATION)),
        early_stopping=_bool_model_param(params.get("early_stopping"), DEFAULT_EARLY_STOPPING),
        random_state=random_state,
    )

The estimator: a histogram-based gradient-boosted regressor, regularized (leaf-size + L2) and early-stopped to resist overfitting.

FORBIDDEN_ZONE_TOKENS = (
    "forward", "post_signal", "post-signal",
    "outcome", "label", "future", "exit",
)
# A predictor is rejected before training if it names a banned
# (projected / full-day / outcome) path, or isn't on the frozen allowlist.
if locator in I12_PIT_BANNED_FEATURE_PATHS:
    raise FeatureSelectionError(
        "I12 PIT-clean feature schema requested a banned projected, "
        f"full-day, or outcome field: {field!r}"
    )
if locator not in I12_PIT_0940_ALLOWED_FEATURE_PATHS:
    raise FeatureSelectionError(
        "I12 PIT candidate feature is not in the frozen inclusion "
        f"allowlist: {field!r}"
    )

Look-ahead made structurally impossible: any feature touching a forward/outcome zone, or absent from the frozen allowlist, raises and halts training, rather than silently inflating a result.

embargo = max(horizon_sessions, embargo_sessions)
train_cutoff = min(test_positions) - embargo
test_identities = {
    _security_identity_key(row)
    for row in examples if row.signal_date in set(block)
}
# Train rows must be strictly earlier than the test block (date purge)
# AND share no security with it (cluster purge), so nothing overlaps.
train_indices = [
    idx
    for idx, row in enumerate(examples)
    if date_to_position[row.signal_date] < train_cutoff
    and _security_identity_key(row) not in test_identities
]

The CV split is both date-purged (train strictly precedes test by the full horizon) and cluster-purged (no security appears on both sides), closing the two ways a time-series backtest usually leaks.

Reproducibility trace

The honest read: the model found measurable rank separation versus naive baselines, but the book still failed the economic bar after spread, slippage, and walk-forward validation. A model that scores better than chance yet loses to cost is not a tradeable edge; the platform is built to report that, not bury it.

A result advances only if every gate agrees

The platform separates validation into two decisions: whether the data is trustworthy, and whether the strategy is tradeable. For the worked example, the data passed; the strategy did not. That distinction is the point.

Group A · Data & corpus gatesPassed

Group B · Feature & leakage gatesPassed

Group C · Economic & promotion gatesFailed

Forward-valid gate evidence

Representative artifacts: i12_pit_stage1_analysis_5snap_3f5ed25.json · i12_pit_0940_walk_forward_threshold_gate_3f5ed25.json · i12_pit_oracle_full_day_volume_labels_5snap_3f5ed25.json.

Five decision timestamps, not one convenient fill

Every candidate was evaluated at five point-in-time decision timestamps: 09:35, 09:40, 09:45, 10:00, and 10:30. The point is not to cherry-pick the best timestamp after the fact; it is to test whether a signal survives realistic entry timing rather than one convenient, assumed fill.

For each row, predictors were computed only from bars available before that decision timestamp. Later snapshots and full-day outcomes were used as labels and diagnostics only, never as features.

Legacy next-open diagnostic