Quant Fund Alternative Data Integration: How Systematic Funds Source, Process, and Trade Satellite, Credit Card, and NLP Signals

The alternative data market has grown from a niche research curiosity to a $7B+ industry in under a decade. Satellite imagery vendors, credit card transaction aggregators, and NLP transcript providers now serve hundreds of institutional clients. The pitch is consistent: access to data that is not in the price, not in the analyst consensus, and not in the standard factor library.

Most systematic hedge funds have heard the pitch. Many have signed vendor contracts. Almost none have a production-grade integration. The gap is not data access — the vendors are commoditized and the evaluation process is straightforward. The gap is the infrastructure layer: ingestion pipelines that clean and normalize heterogeneous vendor formats, a point-in-time database that prevents look-ahead bias in production, and a signal extraction and validation framework that converts raw alt-data into factor-compatible signals without importing the five backtesting biases that make most alt-data pilots look extraordinary in research and flat in live trading.

Our earlier guide to alternative data strategies for institutional investors covers the strategy layer — which alt-data categories carry durable alpha, how signals combine with traditional factor models, and where the edge is eroding. This guide covers the infrastructure layer: what production-grade alt-data integration actually requires, why it is harder than the backtest suggests, and what the build-vs-buy decision looks like for a systematic fund with a $50–300K annual alt-data budget.

This is the third post in the microstructure and data signals sub-cluster, following our guides to market microstructure signals and intraday alpha strategies.

Why Most Alternative Data Programs Fail

Three failure modes account for the majority of alt-data programs that consume significant vendor budget without producing live alpha. Understanding them before writing the first vendor check is the prerequisite for a successful program.

Failure mode 1: The data-before-infrastructure trap. A fund signs a satellite imagery contract, receives vendor credentials, and assigns a quant researcher to evaluate the signal. The researcher downloads weekly parking lot data via the vendor portal, loads it into a Jupyter notebook, constructs a same-store sales proxy, maps it to retail tickers, and produces a backtest showing a Sharpe ratio of 2.1. The research is compelling. The infrastructure does not exist. The data sits in a vendor portal. The notebook runs manually. No ingestion pipeline normalizes the weekly file drops. No point-in-time database stores delivery timestamps. The signal never reaches the production alpha engine. Six months later, the vendor contract renews at $150K and the fund has produced one research note. For the alpha research process framework that governs how a signal moves from hypothesis to production, see our guide to the quantitative alpha research process.

Failure mode 2: The backtest-only signal. The alt-data signal looks extraordinary in the vendor's research dataset — a cleaned, point-in-time-adjusted history with consistent coverage, no delivery gaps, and retroactive panel adjustments already applied. The production pipeline is different. It introduces delivery delays (vendor SLA of 2 business days means the signal available at Tuesday close may not arrive until Thursday morning), normalization errors (the vendor's parquet schema changes quarterly and the fund's ETL job breaks silently), and survivorship bias not present in the vendor's research dataset (tickers that were delisted in 2019 are absent from the historical feed). The live Sharpe is 40% of the backtest. For the backtesting framework that separates research-grade from production-grade signal validation, see our guide to quantitative backtesting best practices.

Failure mode 3: Single-source concentration. A fund builds a working satellite imagery signal for retail foot traffic. It generates consistent positive IC for 18 months. Amazon's Q3 2024 earnings report reveals that the channel-level relationship between parking lot traffic and same-store sales growth has structurally shifted — Amazon Warehouse pickup has absorbed 15% of the brick-and-mortar traffic that previously correlated with revenue. The signal IC drops to near zero. The fund has no diversification across alt-data types. The three-source minimum — satellite imagery, credit card transaction data, and NLP on earnings calls — is not a rule of thumb. It is the minimum required to avoid single-source concentration risk. Any alt-data program with fewer than three uncorrelated sources has idiosyncratic concentration that mirrors single-stock risk in a portfolio without position limits.

The Three Core Alternative Data Categories

The alt-data landscape contains dozens of categories — web scraping, job postings, app usage, shipping data, geopolitical event feeds. The three categories with the deepest institutional adoption, the most mature vendor ecosystem, and the best-documented alpha profiles are satellite imagery, credit card transactions, and NLP on earnings calls. A fund that has productionized all three has a diversified alt-data program. A fund that has productionized one has a research experiment.

2a. Satellite imagery and geospatial data. The signal types range from retail parking lot foot traffic (RS Metrics, Orbital Insight, SpaceKnow) to industrial activity monitoring (oil storage tank levels, shipping container counts, construction site progress rates) and agricultural yield estimation from crop health imagery. The retail parking lot application is the most widely adopted and the most extensively backtested.

Signal construction for retail parking lot follows a consistent pattern: daily car count normalized to the location's square footage, z-scored against the same-day-of-week historical baseline to remove seasonality, converted into a same-store sales growth estimate via a regression on historical earnings data, and ticker-mapped to the parent company. The signal is generated with a 2 to 4 week lag before the earnings announcement — the window in which the foot traffic data has arrived and the forward earnings have not yet been disclosed.

The alpha profile is well-characterized: low frequency (weekly signal update), high Sharpe in backtests (1.8 to 2.5 range), and rapid decay post-analyst adoption. The signal half-life is approximately 2 to 3 years before crowding erodes the IC below the threshold for production deployment. For the signal decay monitoring framework that should govern any satellite data signal, including the IC monitoring thresholds and regime-conditioned expected Sharpe adjustments for post-adoption deployment, see our guide to quantitative signal decay and factor edge management.

The infrastructure requirement is non-trivial: vendor API → normalized geospatial pipeline → location-to-ticker mapping (a parking lot is not a ticker; the mapping from physical location to public company to Bloomberg/Refinitiv identifier requires a dedicated pipeline with ongoing maintenance as store footprints change) → earnings-calendar-gated signal generation. The location-to-ticker mapping is where most satellite data programs fail silently — incorrect or stale mappings generate signal noise that looks like regime-dependent IC degradation when it is actually data quality failure.

2b. Credit card transaction data. Vendors including Earnest Research, Second Measure, Bloomberg Second Measure, and Yodlee aggregate anonymized credit and debit card transaction data from panel providers. The signal types cover consumer spend by merchant category, same-store sales velocity, and wallet share trends — how much of the consumer's discretionary spend is shifting toward or away from a given merchant relative to peers.

Signal construction: weekly spend growth z-scored against a 52-week rolling baseline (removing the trend component to isolate the velocity signal), ticker-mapped at the parent company level (a Starbucks transaction maps to SBUX, not to the franchise operator), signal generated 2 to 3 weeks before earnings, and combined with the analyst consensus revision signal as a confirmation layer. Spend acceleration with upward revision momentum outperforms either signal alone by a documented 15 to 25 bps IC improvement. For the data infrastructure that supports multi-source signal combination, including normalization and ticker mapping at scale, see our guide to quant fund data infrastructure.

The critical bias in credit card data that must be modeled explicitly: panel selection. Credit card panels skew toward certain demographics (higher-income, urban, card-carrying) and geographies (US-centric, underrepresenting cash-dominant markets). The systematic consequence: signals for consumer discretionary stocks outperform signals for consumer staples, and signals for US-listed companies outperform signals for multinational companies with large international revenue exposure. A credit card spend signal for McDonald's systematically understates the portion of revenue that comes from international markets where card penetration is lower.

Infrastructure requirement: vendor data delivery (FTP/SFTP weekly file drops — most credit card vendors do not have real-time APIs; they deliver weekly panel aggregations via scheduled file delivery) → panel stability monitoring (the panel composition changes as vendors add and lose data partnerships; a 15% panel size change week-over-week is a data quality event, not a signal event) → normalization pipeline → ticker mapping → signal generation → production integration.

2c. NLP on earnings calls and SEC filings. The signal types include sentiment scoring on earnings call transcripts (linguistic uncertainty, hedging language frequency, management tone delta vs. the prior four-quarter baseline), 8-K and 10-Q filing complexity metrics (Loughran-McDonald word list applied to identify complexity, uncertainty, and litigation risk language), and analyst question hostility scores (how confrontational are analyst questions vs. the prior two-quarter baseline — a leading indicator of management credibility concerns).

Signal construction: earnings call transcript via transcript API (Refinitiv Transcript API, Bloomberg, or Tegus) → sentence-level sentiment scoring using a domain-specific financial NLP model (Loughran-McDonald lexicon for lexicon-based approaches, or a FinBERT fine-tune for contextual scoring) → delta vs. the prior four-quarter baseline for the same company (absolute tone is less informative than tone deterioration) → combined into a management tone deterioration score → used as a confirmation or veto signal for momentum and quality factors. A momentum long position in a name with a management tone deterioration score exceeding 1.5 standard deviations is a candidate for size reduction. For the machine learning pipeline that supports FinBERT deployment and NLP model versioning, see our guide to machine learning in quantitative finance.

The alpha profile: orthogonal to price-based signals (IC correlation with momentum, value, and quality factors below 0.10), moderate standalone Sharpe (0.8 to 1.2 range), and high value as a confirmation signal — adding 15 to 30 basis points when combined with a multi-factor model that already captures momentum and quality.

Infrastructure requirement: transcript API → NLP scoring pipeline → factor integration layer → real-time ingestion for 8-Ks via SEC EDGAR RSS feed (the SEC publishes a real-time RSS feed for Form 8-K filings that enables NLP scoring within minutes of filing, before the market processes the disclosure).

The Alternative Data Infrastructure Stack

The alt-data infrastructure stack has four layers. Each is necessary. The signal quality in production is limited by the weakest layer — a sophisticated NLP pipeline built on top of an unreliable ingestion layer produces sophisticated noise.

Layer 1: Ingestion layer. Vendor API connectors and file drop monitors, raw data landing zone, and format normalization. The heterogeneity of vendor formats is a practical problem that most funds underestimate: satellite imagery vendors deliver weekly CSV files; credit card vendors deliver parquet with weekly SFT drop; NLP transcript vendors deliver JSON via REST API with per-transcript endpoints. A unified ingestion layer must handle all three formats, detect delivery anomalies (a missing weekly file drop is a data quality event), and normalize to a consistent internal schema without modifying the raw landing zone data.

Delivery delay monitoring is non-negotiable. Most alt-data vendors carry a 1 to 5 business day delivery SLA with material variance. The ingestion layer must log the delivery timestamp for every record — the moment the data arrived at the fund's systems — and flag SLA breaches immediately. A credit card file that arrives 4 days late must be stored with the correct delivery timestamp, not the observation timestamp, or every backtest run against that data is contaminated with implicit look-ahead bias. For the data infrastructure architecture that handles delivery delay monitoring at scale, see our guide to real-time market data infrastructure.

Layer 2: Point-in-time database. The single most important infrastructure investment in an alt-data program. All alt-data must be stored with three timestamps: the observation timestamp (when the underlying event occurred — the parking lot count on Tuesday October 1st), the vendor delivery timestamp (when the vendor published the data — Friday October 4th per their SLA), and the fund's ingestion timestamp (when the data arrived in the fund's systems — Friday October 4th at 3:47pm).

The backtest simulation must only use data that would have been available at the simulated decision point — the fund's ingestion timestamp, not the observation timestamp or the vendor delivery timestamp. Most vendor research datasets are already point-in-time adjusted; production pipelines rarely are, because the infrastructure to track all three timestamps at every delivery event is non-trivial to build and is not part of any standard data warehouse template. This gap — research dataset is PIT-adjusted, production pipeline is not — is the most common cause of backtest-to-live performance degradation in alt-data programs.

Layer 3: Signal extraction layer. Transforms raw alt-data into factor-compatible signals: z-scored, normalized, stationary, and mapped to the instrument universe. This layer must be versioned — every change to signal logic must be auditable, with the version, author, date, and rationale stored in the signal registry. It must also be validated: out-of-sample validation is required before any signal logic change reaches the production alpha engine. A signal that was validated on data through 2021 must be re-validated when signal logic is modified, even if the modification appears minor.

Layer 4: Factor model integration. Alt-data signals are not standalone factors — they enter the alpha engine as additional inputs to the existing factor model, with appropriate half-life weighting and IC validation. The integration architecture treats alt-data signals as a new factor cluster, not a separate strategy. A satellite parking lot signal that predicts retail same-store sales is an additional input to the value factor cluster, not a standalone alpha module with its own position sizing, risk budget, and P&L attribution. Factor model integration ensures that the alt-data signal contributes to the portfolio in proportion to its validated IC, with the half-life weighting that matches its documented signal persistence.

Backtesting Alternative Data: The Five Biases That Make Alt-Data Backtests Unreliable

This section is the highest-value content for a quant researcher evaluating an alt-data program. It explains why 90% of alt-data pilot programs produce compelling backtests and flat live performance. Each bias is a systematic inflation mechanism that makes the backtest look better than the live strategy will perform — not because the data lacks alpha, but because the production pipeline will not replicate the conditions under which the backtest was run.

Bias 1: Look-ahead bias via non-point-in-time data. A fund backtests a credit card signal using the vendor's research dataset. The vendor's dataset has applied retroactive panel adjustments — a methodology change in 2022 revised panel weights going back to 2018, improving coverage accuracy. The research backtest incorporates the adjusted data throughout 2018–2022. The production pipeline receives unadjusted data first; the retroactive adjustment arrives 6 to 12 months later as the vendor publishes a dataset refresh. In live trading from 2022 to 2023, the fund's production pipeline was running on the pre-adjustment data while the backtest was run on the post-adjustment data. The mitigation is explicit: build your own point-in-time database that stores both the original delivery and the revised delivery; never backtest directly from a vendor's research file.

Bias 2: Selection bias in vendor history. Vendors curate their historical datasets toward their strongest-performing signal segments. A satellite imagery vendor that launched in 2018 with 500 US retail locations has expanded to 2,400 locations by 2024. The 2024 research dataset presents the historical data as if the 2,400-location coverage existed in 2018 — the vendor has backfilled location data wherever it can. The 2015 data in the vendor's research file looks like 2024 coverage. It does not. The mitigation: adjust the backtest universe to match the actual historical coverage of the dataset at each point in time. This requires requesting a coverage vintage history from the vendor — how many locations were active in each month of the history. Most vendors can provide this; most funds never ask for it.

Bias 3: Survivorship bias in ticker mapping. Tickers that were acquired, went bankrupt, or delisted are systematically underrepresented in vendor alt-data histories. A credit card vendor's historical dataset contains spend data for Sears, RadioShack, and Pier 1 through the years they were active — but the ticker mapping for delisted securities is often absent or incomplete, and the backtest universe excludes them. This inflates backtest returns by removing the distressed consumer retailers with declining spend trends that would have generated signal losses. CRSP-equivalent delisted security inclusion is required for any alt-data backtest that covers the consumer sector. For the full backtesting bias framework across all signal types, see our guide to quantitative backtesting best practices.

Bias 4: Signal crowding decay not modeled. Alt-data signals have finite half-lives once broadly adopted by the institutional community. A satellite parking lot signal backtest from 2018 includes the 2015–2018 pre-adoption period, when fewer than a dozen institutional funds were licensing satellite data. Live deployment in 2026 is post-adoption: over 200 systematic funds are estimated to license satellite retail foot traffic data from major vendors. The crowded trade mechanics are well-documented — the same signal held by many institutional buyers is absorbed into price faster, reducing the signal's holding period alpha. The mitigation: regime-condition the expected Sharpe by adoption proxy — the number of funds known to license the dataset, available from vendor disclosure statements and prime broker prime brokerage market color. The post-adoption expected Sharpe for a satellite parking lot signal is approximately 40 to 60% of the pre-adoption backtest Sharpe.

Bias 5: Transaction cost underestimation for high-frequency alt-data signals. Credit card data signals update weekly. The backtest often assumes daily rebalancing with academic-level transaction cost assumptions — 5 to 10 basis points round trip for mid-cap names. Production costs for a weekly signal include market impact from the rebalancing trade (Almgren-Chriss square-root model for mid-cap names generates 15 to 30 basis points market impact at 10% ADV position sizes), bid-ask spread (3 to 8 basis points round trip for mid-cap names in 2026), and timing risk from executing a weekly signal across a trading session. Production transaction costs reduce realized Sharpe by 30 to 60% for signals with weekly holding periods. For the TCA framework that calibrates production transaction costs against alt-data signal holding periods, see our guide to quantitative transaction cost analysis.

Signal Validation and Production Deployment

The validation gate between a successful alt-data backtest and production deployment has three stages. Skipping any stage is the standard path to the canonical failure timeline described in Section 6.

Stage 1: Out-of-sample backtest. 40/60 in-sample/out-of-sample split, with the out-of-sample period matching the expected live deployment regime. For a fund deploying a signal in 2026, the out-of-sample period should be 2022 to present — the post-rate-hike, post-COVID normalization regime that will be the operating environment for live deployment. A signal that produces a Sharpe of 1.8 in-sample and 0.7 out-of-sample is failing the validation gate regardless of the in-sample statistic. The out-of-sample statistic is the signal's production forecast. For the out-of-sample validation methodology that avoids false OOS independence (using the same optimization universe for both IS and OOS), see our guide to the quantitative alpha research process.

Stage 2: Paper trading period. Eight to twelve weeks of live signal generation without position-taking. The paper trading period tests the production pipeline, not the signal hypothesis. During this period, the fund monitors: real-time delivery delay against the vendor's SLA (are files arriving within the documented window, and what is the actual P50 and P99 delay?), panel stability for credit card data (is the weekly panel size stable within 15%?), and signal IC calculation against forward returns (is the live-production signal generating the IC the backtest predicted, or is there a systematic degradation from normalization or ticker mapping errors?).

Stage 3: Small-size live deployment. 5 to 10% of target notional for the first 6 months. Full TCA attribution, weekly IC tracking, and a pre-specified decision rule for scaling up (rolling 12-week IC z-score above 0.5 for 8 consecutive weeks) or abandoning the signal (rolling 12-week IC z-score below 0.3 for 4 consecutive weeks with no identifiable regime explanation).

Ongoing monitoring requirements after production deployment: a panel size stability alert for credit card data triggering review when week-over-week panel size change exceeds 15%; a delivery delay SLA breach alert for any vendor with more than 2 consecutive SLA misses; a signal IC z-score decay alarm with rolling 12-week IC z-score below 0.5 as the review trigger; and a vendor contract renewal trigger that validates the renewal decision against the current live signal IC, not the historical backtest IC. For the model risk framework that governs ongoing alt-data signal monitoring, including formal demotion triggers and the champion-challenger testing infrastructure, see our guide to quant fund model risk management.

Build vs. Buy: The Real Infrastructure Decision

The build-vs-buy decision for alt-data is not about the data subscriptions — those are purchases. The decision is about the infrastructure layer, and it has a clear answer: most of the infrastructure must be built, even by funds that are otherwise platform-native.

What to buy. Alt-data vendor subscriptions. Start with one dataset per category: satellite → RS Metrics or Orbital Insight for retail foot traffic; credit card → Second Measure (now Bloomberg Second Measure) for consumer spend; NLP → Refinitiv Transcript API or FinBERT open-source for earnings call scoring. Budget: $50 to $300K per year for a three-dataset program at a $200M to $1B AUM fund. The vendor evaluation framework — coverage vintage history request, panel stability documentation, delivery SLA terms, MNPI legal memo for alt-data inputs — is covered in our guide to hedge fund technology for CTOs. The budget-allocation framework for a three-dataset program relative to total technology spend is covered in our guide for CFOs and COOs.

What to build. Point-in-time database schema: non-negotiable, cannot be purchased from a vendor because it must integrate with the fund's own ingestion timestamps and delivery delay monitoring infrastructure. Signal extraction layer: vendor-specific, because the normalization logic for RS Metrics satellite CSV is different from the normalization logic for Second Measure parquet. Ticker mapping infrastructure: location-to-ticker for satellite, company-to-ticker with multi-subsidiary support for credit card (a Yum! Brands transaction can map to KFC, Pizza Hut, or Taco Bell — the parent ticker mapping must handle all three). IC monitoring dashboard with the rolling 12-week z-score alarms and vendor SLA breach notifications. For the full data infrastructure architecture — including PIT database patterns by AUM tier and the vendor evaluation framework for alt-data providers — see our guide to quant fund data infrastructure.

What most funds do wrong. Buy three datasets. Backtest in notebooks. Never build the point-in-time database. Run a 6-month pilot that shows a 2.1 Sharpe. Deploy to production. Underperform by 40 basis points due to delivery delays and normalization errors that were absent from the vendor's research dataset. Cancel the vendor contract 9 months after live deployment. Conclude that alternative data does not work for systematic funds.

The failure timeline is consistent: Q1 buy data → Q2 backtest with vendor research file (looks excellent) → Q3 deploy to production without PIT database → Q4 underperform by 30 to 50 bps → Q1 cancel vendor contract. The infrastructure layer takes longer to build than the backtest. A fund that commits to building the PIT database first — before signing any vendor contract — runs the process in the right order. Four months of infrastructure work before first data purchase is not a delay; it is the prerequisite for the program producing live alpha rather than a research curiosity.

AlphaEdge AI. AlphaEdge AI provides the pre-built alt-data infrastructure layer: ingestion connectors for satellite CSV, credit card parquet, and NLP JSON delivery formats; point-in-time storage with three-timestamp schema (observation, vendor delivery, fund ingestion) and delivery delay tracking; signal extraction templates for all three alt-data categories with version control and validation gates; IC monitoring dashboard with rolling 12-week z-score alarms; and factor model integration layer that maps alt-data signals into the existing alpha engine as a new factor cluster rather than a standalone strategy. For the full technology stack context — where alt-data integration sits in the systematic fund infrastructure roadmap — see our guide to the quant hedge fund technology stack in 2026. For the crowding risk layer that sits alongside alternative data signals — how to detect and systematically reduce crowded factor exposure before an unwind event — see our guide to quant fund factor crowding risk management.

Alternative Data Integration: 20-Point Implementation Checklist

Use this checklist to assess your current alt-data infrastructure and identify the highest-priority gaps before the next vendor contract renewal or new data category evaluation.

Infrastructure Foundation (5)

• Point-in-time database schema defined and implemented before any data purchase — three-timestamp schema (observation, vendor delivery, fund ingestion) deployed and validated against at least one historical vendor delivery cycle
• Delivery delay monitoring with SLA breach alerts — every vendor file delivery logged with ingest timestamp; alert fires within 4 hours of SLA window closing without confirmed delivery
• Signal versioning infrastructure deployed — all signal logic changes are auditable with version number, author, date, rationale, and validation status stored in signal registry
• Factor model integration layer configured — alt-data signals enter the alpha engine as a new factor cluster with half-life weighting and IC-based allocation, not as standalone strategies with independent position sizing
• Three-source minimum enforced — no alt-data program considers itself operational with fewer than satellite, credit card, and NLP sources in production or validated paper trading

Data Source Configuration (5)

• Panel stability monitoring for credit card data — weekly panel size check with 15% week-over-week threshold triggering data quality review before signal generation runs
• Location-to-ticker mapping pipeline for satellite data — mapping table maintained with quarterly refresh, covering store openings, closures, and parent company changes; stale mapping alerts fire when location count diverges from vendor metadata by more than 5%
• Company-to-ticker mapping with multi-subsidiary support — credit card and NLP signals mapped to ultimate parent at Bloomberg/Refinitiv identifier level, with subsidiary-to-parent table maintained for multi-brand consumer companies
• FinBERT or Loughran-McDonald NLP pipeline deployed for transcript scoring — model version logged for every scoring run; delta vs. prior 4-quarter baseline computed per issuer per quarter
• SEC EDGAR RSS feed integrated for real-time 8-K ingestion — NLP scoring fires within 10 minutes of 8-K publication; filing timestamp logged for PIT compliance

Backtesting Integrity (5)

• Out-of-sample backtest with 40/60 split — no signal validated exclusively on vendor research files; OOS period post-2022 for all signals deployed in 2026
• Survivorship bias correction enforced — delisted, acquired, and bankrupt securities included in backtest universe with CRSP-equivalent delisted security mapping for every sector backtest
• Selection bias adjustment applied — historical coverage mapped by year from vendor vintage history; backtest universe restricted to actual coverage at each point in time, not current coverage projected backward
• Signal crowding proxy built — adoption proxy (known licensee count estimate) maintained for each dataset; expected post-adoption Sharpe calculated and documented alongside full-history backtest Sharpe
• Transaction cost model calibrated to actual holding period — weekly signal round-trip costs modeled at Almgren-Chriss market impact for 10% ADV positions, not daily rebalancing academic assumptions; realized Sharpe forecast reflects production TC model

Validation & Deployment (5)

• Paper trading period minimum 8 weeks before production — live signal generation with real delivery latency tracked, panel stability confirmed, and IC calculated against forward returns before first position taken
• Small-size live deployment — 5–10% of target notional for first 6 months, with pre-specified scale-up criteria (rolling 12-week IC z-score above 0.5 for 8 consecutive weeks) and abandonment criteria (IC z-score below 0.3 for 4 consecutive weeks)
• IC monitoring active — rolling 12-week IC z-score computed per signal per week; 0.5 threshold triggers review; 0.3 threshold triggers demotion process under model risk governance framework
• Panel size stability alert deployed — 15% week-over-week threshold active for all credit card data sources; alert suppresses signal generation until data quality review is completed and signed off
• Vendor renewal decision process documented — renewal evaluated against rolling 12-week live IC and TCA-adjusted realized Sharpe, not historical backtest; renewal decision memo stored in model risk governance documentation