Dataset card (CC BY 4.0)

README.md

@@ -0,0 +1,119 @@
+---
+license: cc-by-4.0
+tags:
+- Steam
+- video-games
+- game-development
+- revenue-estimation
+- indie-games
+- causal-inference
+- game-analytics
+- Boxleiter
+- open-data
+pretty_name: Steam Launch Lever Panel — 2026
+---
+
+# Steam Launch Lever Panel — 2026
+
+**10,403 Steam launches with public-source-derived week-1 revenue, plus the storefront features used to estimate causal lever effects (demos, co-op/multiplayer, localization).**
+
+License: **CC BY 4.0** · Compiled by [SteamForecast](https://steamforecast.app)
+Canonical analysis: **https://steamforecast.app/reports/steam-launch-lever-benchmarks**
+Methodology: **https://steamforecast.app/methodology**
+
+---
+
+## ⚠️ Read this before using the revenue column
+
+**This panel is survivorship-biased, and the bias is large.** Games that sold almost nothing leave little owner-count or review trace, so no revenue can be derived for them and they silently drop out of the denominator. What remains is the subset of Steam launches with *measurable* revenue.
+
+The practical consequence: **the median `y_week1_cents` in this file is ~$189,600, and that is emphatically NOT "the median Steam game."** The median Steam release earns a small fraction of that. This file's median describes the median *measurable* launch, which is a very different population.
+
+Do not quote an absolute median-revenue figure from this dataset without carrying that caveat in the same breath. The dataset is built for **relative** comparisons — how much a lever moves revenue, holding other features constant — which is where survivorship bias largely cancels. It is not built to answer "what does a typical Steam game make."
+
+Longer treatment of why most published Steam benchmarks share this flaw: https://steamforecast.app/reports/steam-wishlist-benchmarks-reality-check (that report uses a differently-filtered n=10,416 for its own question — the two counts are both correct for their respective filters, not a discrepancy).
+
+---
+
+## Files
+
+| File | Rows | Notes |
+|---|---:|---|
+| `steam-launch-lever-panel-2026.csv` | 10,403 | UTF-8, header row |
+| `steam-launch-lever-panel-2026.parquet` | 10,403 | identical content, typed |
+
+## Columns (19)
+
+**Identity**
+| Column | Type | Description |
+|---|---|---|
+| `appid` | int | Steam application ID |
+| `name` | str | Store title at extraction |
+| `release_date` | date | Steam release date (1997-06-30 → 2026-05-01) |
+| `release_year` | int | Convenience split of `release_date` |
+
+**Outcome (derived — see provenance)**
+| Column | Type | Description |
+|---|---|---|
+| `y_week1_cents` | int | Estimated week-1 gross revenue, USD cents |
+| `y_log_week1` | float | log10 of the above; the modeling target |
+
+**Treatments (the levers under study)**
+| Column | Type | Description |
+|---|---|---|
+| `T_has_demo` | 0/1 | Demo shipped at launch (13.7% of rows) |
+| `T_supports_coop` | 0/1 | Co-op or multiplayer (14.8%) |
+| `T_languages_high` | 0/1 | ≥10 supported languages (23.8%) |
+| `T_price_bin` | str | `0_under_5` / `1_5_to_10` / `2_10_to_20` / `3_20_to_30` / `4_30_plus` |
+| `T_release_month` | int | 1–12, seasonality control |
+
+**Controls**
+| Column | Type | Description |
+|---|---|---|
+| `recommendations_total` | int | Total reviews at extraction |
+| `num_genres` | int | Store genre count |
+| `genre_cluster` | str | `action_arcade` / `strategy_sim` / `rpg_adventure` / `casual_puzzle` / `other` |
+| `has_metacritic_score` | 0/1 | Metacritic present (26.1%) |
+| `is_mature_content` | 0/1 | Mature flag (0.3%) |
+| `platform_mac` / `platform_linux` | 0/1 | Additional platform support (35.8% / 25.2%) |
+
+**Provenance**
+| Column | Type | Description |
+|---|---|---|
+| `source_quality` | str | How the revenue estimate was derived — **B / C / D**, see below |
+
+No nulls, and no all-constant columns — two fields (`follower_count`, store `tags`) existed in our internal panel but were never populated for this corpus, so they are **excluded** rather than shipped as columns of zeros. Consequence worth knowing for replication: the published lever estimates do **not** control for community-follower counts or store tags.
+
+## How `y_week1_cents` is derived — and its limits
+
+These are **estimates from public signals, not reported sales.** No sales figures were licensed, purchased, or obtained from any storefront partner or private analytics feed. The chain is:
+
+1. **Lifetime owners** estimated by one of two public paths, recorded per row in `source_quality`:
+
+   | Tier | n | Method |
+   |---|---:|---|
+   | **C** | 7,652 | SteamSpy `owners` range → geometric mean |
+   | **B** | 2,734 | Boxleiter (reviews × NB multiplier) **capped** by the SteamSpy upper bound — the most constrained path |
+   | **D** | 17 | Boxleiter alone (`reviews × 63`, the 2025 multiplier), no SteamSpy cross-check — weakest tier |
+
+2. **Lifetime revenue** = estimated owners × price.
+3. **Week-1 share** applied by ownership tier, calibrated from public postmortems and industry write-ups (How To Market A Game, GameDiscoverCo, Steam Discovery Update notes): AAA (≥1M owners) 0.12 · mid-major (200k–999k) 0.20 · mid-tier (50k–199k) 0.30 · small indie (<50k) 0.40. A flat ratio badly over-derives mega-hit targets, which is why the tiering exists.
+
+**Known limitations, stated plainly:**
+- Every revenue figure carries compounding error from three estimated stages. Treat individual rows as indicative, not authoritative. **If a specific game's number looks wrong to someone who knows the real figure, it may well be** — that is inherent to review-proxy estimation, and `source_quality` tells you how much to trust each row.
+- Gross revenue before Steam's cut, refunds, regional pricing, discounts, taxes, and chargebacks.
+- Free-to-play titles are excluded entirely (`is_free = FALSE`); revenue proxies do not work for them.
+- Only launches with derivable revenue appear — the survivorship point above.
+- `release_date` spans 1997–2026; older entries reflect a very different Steam.
+
+## Filters applied
+
+`week1_revenue > 0` · `price > 0` · not free-to-play · `release_date` present.
+
+## Suggested citation
+
+> SteamForecast (2026). *Steam Launch Lever Panel — 2026* (n=10,403). CC BY 4.0. https://steamforecast.app/reports/steam-launch-lever-benchmarks
+
+## What is NOT in this dataset
+
+Wishlist counts are **not** included and never will be in this file. Pre-launch wishlist data is not public, and our wishlist-aware modeling is trained on a separately licensed feed that we cannot redistribute. Every column here derives from public Steam storefront data, public community data, SteamSpy, or public review counts.