App Download And Usage Data For Consumer Companies

When to Use

Use this skill when processing mobile application alternative data (sourced from vendors such as Sensor Tower, data.ai, Apptopia, or Similarweb) to build fundamental engagement signals for consumer-facing public companies. While raw app downloads are often reported as "vanity metrics" by companies, the actual Daily Active Users (DAU) and Monthly Active Users (MAU) dictate long-term revenue viability. This engine calculates the "Stickiness Ratio" (DAU/MAU) and flags dangerous divergence between high marketing-driven downloads and failing user retention (the "Leaky Bucket" syndrome).

Applicable scenarios:

• Forecasting revenue durability for consumer-tech, gaming, ride-share, food-delivery, and streaming issuers ahead of earnings.
• Constructing long/short alt-data baskets (overweight is_world_class, underweight churn_risk_warning).
• Validating management commentary on "record downloads" against underlying engagement.

When NOT to Use

Do not use this skill when:

• You have not yet performed vendor diligence and MNPI/MAR compliance review on the data source. App Annie Inc. was the subject of the SEC's first securities-fraud enforcement against an alternative data provider (Release No. 34-92975, Sept. 14, 2021): it promised the app companies supplying its data that the data would be used only in aggregated, anonymised form, then used it in non-aggregated form to adjust its model estimates while assuring subscribers those estimates were generated consistently with the consents obtained. It is a provenance case, not a case about public data being per se MNPI. See insider-trading-controls-for-alternative-data-usage.
• The issuer's revenue is not materially driven by app engagement (e.g., pure B2B, hardware-only, or pre-product companies with negligible MAU).
• You need raw point-in-time alignment. This engine consumes already-PIT-aligned AppUsageDataPoints; perform the publication-lag shift upstream via alternative-data-feature-integration first.
• You need daily panel spend / transaction signals — use credit-card-transaction-data-signal-construction instead.
• You have not calibrated the stickiness thresholds to the issuer's app category. The 50% / 20% defaults are consumer-social rules of thumb; published 2025 benchmarks put healthy e-commerce (~20-23%) and insurance (~16-27%) apps at or below the 20% "low engagement" line. Running the defaults over a weekly-cadence category produces sector-wide false churn warnings.

Prerequisites

• Python 3.10+.
• A vendor feed providing per-ticker, per-date downloads, dau, and mau estimates.
• Completed vendor diligence: panel methodology documentation, data-licensing terms permitting investment use, and a documented MNPI/MAR compliance sign-off (see alternative-data-vendor-due-diligence-checklist).
• Point-in-time alignment of the feed (event date shifted by the vendor's publication lag, typically 1-7 days).

Workflow

1. Vendor Ingestion & Diligence: Acquire panel data (Downloads, DAU, MAU) per ticker. Confirm the vendor's panel composition, extrapolation model, and licensing terms permit investment use. Document the publication lag. Decision point: before nominating a second vendor for cross-validation, verify it is independent in ownership and panel — Sensor Tower acquired data.ai (formerly App Annie) in March 2024 and merged its panel, so reconciling those two is not corroboration.
2. Point-In-Time Alignment: Shift event dates forward by the vendor's publication lag via alternative-data-feature-integration so signals are only usable on the date the data became available (eliminates look-ahead bias).
3. Window Alignment: Aggregate downloads over the window high_acquisition_fraction is calibrated against — the 0.10 default assumes a trailing 30-day sum, matching MAU's window. Decision point: if you pass raw single-day downloads against the default, churn_risk_warning will never fire; that is a silent false negative, not an error. Aggregate to 30 days or recalibrate the fraction to a daily scale.
4. Threshold Calibration: Set world_class_threshold and low_stickiness_threshold from a category peer cohort's observed distribution rather than the defaults, and record the values and cohort with the signal. See references/standards.md -> "Threshold provenance and calibration".
5. Ingest Metrics: Load PIT-aligned, window-aligned data into AppUsageDataPoint objects (frozen; ticker, date, downloads, dau, mau). Counts must be finite real numbers; NaN/inf raise ValueError and non-numeric values (None, strings, bool) raise TypeError, so a panel gap can never be scored as engagement.
6. Calculate Stickiness: AppUsageSignalEngine.process() derives stickiness_ratio = DAU / MAU. If DAU > MAU (impossible in a genuine panel), DAU is clamped to MAU without mutating the input and the event is logged as a data anomaly.
7. Analyze Divergence: The engine compares download velocity against stickiness. With the defaults, downloads >= 10% of MAU while stickiness < 20% emits churn_risk_warning=True.
8. Signal Generation: Output an AppUsageSignal classifying the issuer's user base as world-class engagement, leaky-bucket churn risk, or average.
9. Decision Points:
   • Overweight equities flagged is_world_class.
   • Underweight/short equities flagged churn_risk_warning — but first confirm the thresholds were category-calibrated; an uncalibrated warning on a weekly-cadence app is most likely a threshold artefact, not a churn finding.
   • Hold equities with average engagement; treat as noise unless combined with other alt-data signals.

Full procedure: see references/workflows.md. Standards reference: see references/standards.md. Printable pre-flight checklist: see assets/checklist.md.

Common Pitfalls

• Confusing Downloads with Growth: Equating a spike in app downloads (often driven by expensive ad campaigns) with revenue growth. If DAU/MAU is well below the issuer's category norm, newly acquired users are churning out immediately.
• Applying Social-App Thresholds Across Categories: The 50% world-class and 20% low-engagement defaults describe habitual consumer-social apps. Published 2025 benchmarks put whole healthy verticals at or under 20% (e-commerce ~20-23%, insurance ~16-27%). A food-delivery or airline app is designed for weekly use — scoring it against a messaging-app bar produces a sector-wide short basket built on a calibration error, not on churn.
• Mismatching the Downloads Window: downloads is a flow, mau a 30-day stock. Comparing single-day downloads against the 30-day-calibrated high_acquisition_fraction=0.10 makes the leaky-bucket condition effectively unreachable — a real app rarely acquires 10% of its monthly actives in one day. The failure is silent: no warning ever fires, and nothing raises.
• Assuming Two Vendor Brands Are Two Vendors: Sensor Tower acquired data.ai (formerly App Annie) in March 2024 and folded its panel in. Reconciling a Sensor Tower estimate against a data.ai estimate will not surface a shared panel bias or a common extrapolation error, however independent the two dashboards look.
• Ignoring Point-in-Time (PIT): Backtesting on the event date rather than the vendor publication date introduces look-ahead bias. Always shift by the publication lag upstream (see alternative-data-feature-integration).
• Skipping MNPI / Vendor Diligence: Consuming app-usage estimates without confirming the vendor's derivation methodology and MNPI controls risks repeating the App Annie (SEC 34-92975) failure mode — trading on data whose provenance and compliance posture were misrepresented.
• Trusting DAU > MAU: DAU can never exceed MAU. A panel reporting otherwise is a data-quality defect; the engine clamps and logs it, but recurring occurrences indicate a vendor panel problem requiring escalation.
• Overreading Cumulative Downloads: Cumulative downloads are a poor standalone proxy for enterprise value; always pair with engagement (stickiness) and acquisition-cost context. Treat the strength of that relationship as something to measure on your own panel — no quantified correlation is asserted here.
• Silent NaN From Panel Gaps: Vendor exports encode gaps as NaN. Every threshold comparison against NaN is False, so without validation a missing observation emits a well-formed "Average engagement" signal with stickiness_ratio=nan and poisons any downstream aggregate. The engine now rejects non-finite counts outright — never paper over the rejection by substituting 0, which is a real reading meaning "no users".
• Stale Data: App-usage panels update with a lag. Monitoring freshness (last received event date per ticker) is required; a stale feed silently degrades signal quality.

Verification

• Run python -m unittest discover -s skills/app-download-and-usage-data-for-consumer-companies/scripts and confirm all 30 tests pass (covers world-class stickiness, leaky-bucket churn, threshold boundaries, DAU>MAU clamping without input mutation, non-finite and non-numeric counts, fractional vendor estimates, invalid inputs, custom config, and batch processing).
• Manual check: construct an AppUsageDataPoint with dau=6000000, mau=10000000 (stickiness 60%) and confirm is_world_class=True; construct one with downloads=200000, dau=150000, mau=1000000 (stickiness 15%, high acquisition) and confirm churn_risk_warning=True and "LEAKY BUCKET" in the summary.
• Confirm process() does not mutate its input when DAU > MAU (regression test test_dau_exceeds_mau_anomaly).
• Confirm a NaN count raises rather than producing a signal: AppUsageDataPoint(ticker="X", date=..., downloads=1000, dau=float("nan"), mau=1000) must raise ValueError (regression test test_nan_dau_raises_rather_than_emitting_nan_signal).
• Confirm the thresholds in use were calibrated to the issuer's app category, and that the recorded downloads window matches the window high_acquisition_fraction was calibrated against.
• Confirm a documented MNPI/vendor-diligence sign-off exists before promoting any signal to live trading.

Related Skills

• alternative-data-feature-integration
• insider-trading-controls-for-alternative-data-usage
• alternative-data-vendor-due-diligence-checklist
• backtesting-alt-data-strategies-with-realistic-availability-lag
• web-scraped-sentiment-data-pipeline