Point-in-Time Alternative Data Ingestion, Entity Resolution &
Survivorship-Free Panel Construction.

Three Ways Alternative Data Backtests Accidentally Cheat

Almost every signal looks great in a naive backtest and then falls apart when real money is on the line. When you audit why alternative data strategies fail in production, the culprit is almost never the math—it is data contamination. Three specific leaks account for most of that gap:

1. The 4:15 PM Filing Trap & Restatement Overwrites: If you align quarterly numbers by fiscal period-end (fp), you are trading on information 30 to 90 days before it was public. Even using the calendar filingDate introduces a full day of lookahead bias whenever a company files its 10-K, 10-Q, 8-K, or Form 4 after the 16:00 ET closing bell (say, at 16:15 ET), because a daily backtest assumes you bought at the 16:00 ET close—15 minutes before the filing hit the wire. Worse, standard data vendors routinely overwrite original quarterly numbers when a company files a restatement months later.

2. Mutable Vendor Timestamps: Many public and commercial data feeds only give you a single last_updated date instead of an immutable event log. If a public registry updates a minor administrative field six months after an event, pulling that record today makes it look as if the original catalyst was missing—or arrived on a completely different date.

3. LLM Pre-Training Contamination: If you ask a commercial language model (trained on internet data through 2025) to read a 2023 earnings release and score whether the stock looks bullish, the model already knows how the stock performed in 2023 and 2024. You aren’t testing a signal; you are testing the model’s memory.

Down-to-the-Second SEC acceptanceDateTime Alignment & Feed Quarantine

To make sure our system only sees what a trader sitting at a desk could see at that exact second, our ingestion layer enforces a strict point-in-time filtration \(\mathcal{F}_t\). Every structured XBRL line item in SEC EDGAR companyfacts (revenue, operating cash flow, working-capital accruals, share counts, and ASC-280 customer footnotes) is joined by accession number (accn) to its EDGAR submissions header to recover the exact down-to-the-second SEC server timestamp \(\tau_{\text{accept}}\) (acceptanceDateTime):

That timestamp routes every disclosure into one of three execution windows: (i) filings accepted before the opening bell (\(\tau_{\text{accept}} < \text{09:30:00 ET}\)) can trade at the same-day \(\operatorname{Open}(T_0)\); (ii) regular-session filings (\(\text{09:30:00 ET} \le \tau_{\text{accept}} < \text{16:00:00 ET}\)) are benchmarked against live arrival mid-quotes (\(\operatorname{ArrivalMid}\)) in our intraday daemon and pushed conservatively to \(\operatorname{Open}(T+1)\) in daily batch replay; and (iii) after-hours filings (\(\tau_{\text{accept}} \ge \text{16:00:00 ET}\)) are pushed strictly to the next trading day’s \(\operatorname{Open}(T+1)\). If a company later files an amendment (10-K/A or 10-Q/A) at \(\tau_{\text{amend}} > \tau_{\text{accept}}\), the historical ledger keeps the original unrevised number over \([\tau_{\text{accept}},\, \tau_{\text{amend}})\) and reveals the amended figure only from \(\tau_{\text{amend}}\) onward.

Quarantining Feeds That Can’t Prove Their History: We apply a simple rule to every data feed: if we cannot verify its historical timestamps down to the minute from an immutable primary archive, we do not allow it to promote itself in a backtest. Scouts backed by immutable primary filing timestamps (SEC EDGAR acceptanceDateTime, Form 4 XML ownership disclosures, 13F-HR institutional holdings, Federal Register publication timestamps, and exchange OHLCV bars) are eligible for historical walk-forward cross-validation. Any scout that ingests a feed with mutable vendor watermarks or rolling short-window APIs is quarantined to `$0` shadow execution, where our own servers log the exact arrival timestamp in real time until the scout earns a live track record out-of-sample.

Point-in-Time Entity Resolution: 10-Digit CIK & OpenFIGI Mapping

Ticker symbols are neither unique nor permanent across multi-year equity panels: bankrupt issuers relinquish tickers that are subsequently reassigned to unrelated listings, and corporate spin-offs or mergers alter symbol mappings mid-sample. Joining alternative disclosures to price series by raw ticker string creates silent cross-entity contamination.

Starling Quantitative resolves every corporate filing, regulatory action, and supply-chain edge at ingestion time to a canonical 10-digit zero-padded SEC Central Index Key (CIK) and a point-in-time composite OpenFIGI identifier before executing cross-sectional joins. Cross-entity supply-chain graphs extracted from Form 10-K Item 1 and ASC-280 major-customer footnotes link upstream suppliers \(i\) to downstream principals \(j\) via verified (CIK_supplier, CIK_principal, tau_accept) edges, ensuring that customer-distress or inventory-glut propagation signals reflect only relationships publicly disclosed prior to evaluation date \(t\).

Survivorship-Free Historical Universe Construction & Shumway Delisting Haircuts

Evaluating fundamental or event-driven strategies across today’s surviving index members systematically truncates left-tail bankruptcy and distressed-acquisition outcomes. To eliminate survivorship bias, Starling’s historical validation panel is constructed from the union of historical SEC EDGAR registrant snapshots and broad-market constituent histories from January 2019 through November 2026 (comprising 505 trading days of rolling estimation warmup prior to out-of-sample evaluation):

Consistent with Shumway (1997), whenever a constituent experiences an unscheduled performance-related delisting without an executable terminal closing print, the backtest engine injects an explicit \(-30.0\%\) delisting return penalty into forward return attribution, preventing distressed value traps from exiting historical simulations at artificially benign last-traded prices.

Deterministic Feature Extraction & Dual Institutional Delivery Schemas

Every Stage-1 domain scout in Starling Quantitative is implemented in 100% deterministic quantitative Python with zero generative LLM calls during primary screening. Given identical point-in-time filing inputs \(\mathcal{D}_{i,t}\), every scout produces bit-for-bit identical feature outputs across historical replay and live production.

Because institutional systematic desks differ in whether they consume raw alternative feature vectors inside their own portfolio optimizer or finished risk-gated trade signals, Starling exposes two distinct machine-readable schemas:

1. Feed Tier A — Pre-Quorum Point-in-Time Feature Stream (ScoutPitch + metric_evidence): Emits un-netted, scout-level records containing top-level ScoutPitch fields (pitch_id, scout_name, target_ticker, instrument_to_buy, action, holding_period_days, stop_loss_pct, take_profit_pct, metric_evidence, and UTC created_at joined to 10-digit SEC CIK and OpenFIGI) alongside standardized inner metric_evidence keys (metric_name, metric_value, norm_mode, polarity, as_of_ts, panel_n_common, feature_domain, pillar_id, half_life_days, z_event \(z_{k,i,t}\), and post_accept_resid_move_bps). Crucially, Tier A preserves the exact target_ticker on both positive and negative (bearish) corporate catalysts prior to any downstream long-only retail ETF substitution, allowing institutional long/short equity desks to harvest single-name short alpha directly from accounting-accrual, dilution, and governance-deterioration disclosures.

2. Feed Tier B — Post-Quorum Calibrated Consensus Stream (MasterDecision): Emits the post-calibration, post-adversarial-veto signal vector after 4-pillar concave pooling, 6-factor Barra residualization, and volatility-parity sizing (detailed in Research Note RN-02 and Research Note RN-03).