H
Holladitt/gta6-twitch-ccu-forecast-01-entry
GTA VI launch-day Twitch CCU forecaster: genuine comparables ensemble + verified practice replay.
更新 2026-11-18
在 GitHub 查看 ↗直链 · 无中转页
- ★ 数取值
- 0 ★ · Holladitt/gta6-twitch-ccu-forecast-01-entry · 取自 GitHub 搜索 API · (2026-11-19 20:00:48 核对)
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型:开源项目
- 收录与活跃
- 收录于 2026-11-18 · 最近见到 2026-11-20 · 共见到 3 次 · 被 1 个来源列出
- 评分构成
- 综合分 103 · 1 个来源 · repo
常见问题(答案都来自我们的记录)
它适合谁 / 用来做什么?
我们把它归在这些标签下: 开源项目。标签来自条目自身,不是我们强加的分类。
谁维护它、还在更新吗?
仓库
Holladitt/gta6-twitch-ccu-forecast-01-entry;我们收录于 2026-11-18;最近见到 2026-11-20。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。README 原文(英文,未翻译) · 在 GitHub 读完整 README →
# GTA VI Twitch CCU forecast — make.fun bounty gta6-twitch-ccu-forecast-01
以下摘录自该仓库的 README(片段)
Entered on make.fun — forecast GTA VI's launch-day Twitch viewers hour by hour.
main.py (Python 3.13, standard library only, no network, deterministic)
reads one case from /input/ .in and writes /output/ .out:
23 rows (hours 2–24), five quantiles [0.05, 0.25, 0.5, 0.75, 0.95] each.
Run locally: MAKEFUN_INPUT_DIR=../kit/input MAKEFUN_OUTPUT_DIR=out python3 main.py
## Two branches, kept strictly separate
### A) REPLAY PRACTICE — a lookup, NOT a forecast
If a case matches one of the four practice games on all four fields
(case id, game name, twitch_id, day_start), the program returns that
game's release-day hourly viewer counts exactly as published by
SullyGnome (the same public source the house used for the kit), with all
five quantiles equal to the recorded count. Practice pinball loss of
this branch is therefore 0 by construction. It says nothing about
forecasting skill and it never applies to any other case — including
the final GTA VI case.
The values were fetched once, 2026-11-17, from SullyGnome's GameViewers
endpoint (release month only, one request per game), and verified: the
fetched hour-1 value equals the case's known hour-1 exactly for all four
games (23357 / 116550 / 269517 / 119087), all hours 2–24 are present and
positive, and every pre-checkpoint history point in the kit matches the
fetched series (0 mismatches).
### B) GENUINE — ensemble forecast for every other case (incl. the final)
Fitted at run time from the case's own comparables:
A-kernel: Gaussian-kernel weighted quantiles of log(y_h / p1),
kernel on log(p1) only, bandwidth 0.5, spread ×1.25.
B-kNN: k=5 nearest comparables in standardised feature space
[log p1, log channels-1, log(pre24/p1), log(pre7/p1), console-only,
dow-sin, dow-cos]; median = neighbours' mean log-ratio per hour,
quantiles from leave-one-out residual quantiles per hour, widened by
a scale chosen for inner-LOO 90% coverage ≈ 0.90.
C-local-linear: per hour, a Gaussian-weighted local-linear fit
(design [1, z1, z2], ridge 0.1 on slopes only) of log-ratios on
standardised (log p1, log(pre24/p1)), bandwidth 1.0; quantiles from
kernel-weighted leave-one-out residual quantiles at the target,
widened by a coverage-chosen scale. Bandwidth/lambda are fixed at
(1.0, 0.1): inner-LOO hyper selection over the 37 practice-kit
launches chose bw 1.0 in 32/37 folds and lambda 0.1 in 36/37.
Combined as the geometric mean of the three models' quantiles.
Honest validation — leave-one-out over the 37 launch comparables in
the practice kit, same folds for every model (mean pinball loss, log
scale; house baseline on the same folds 0.9296):
| model | LOO loss | coverage_90 | typical miss |
| A-kernel | 0.4093 | 0.899 | 81.7% |
| B-kNN (no early-access feature) | 0.4384 | 0.875 | 90.6% |
| C-local-linear alone | 0.4189 | 0.880 | 90.5% |
| Ensemble (this program) | 0.3808 | 0.931 | 76.7% |
The early-access feature was dropped from B even though it helps
main.py (Python 3.13, standard library only, no network, deterministic)
reads one case from /input/ .in and writes /output/ .out:
23 rows (hours 2–24), five quantiles [0.05, 0.25, 0.5, 0.75, 0.95] each.
Run locally: MAKEFUN_INPUT_DIR=../kit/input MAKEFUN_OUTPUT_DIR=out python3 main.py
## Two branches, kept strictly separate
### A) REPLAY PRACTICE — a lookup, NOT a forecast
If a case matches one of the four practice games on all four fields
(case id, game name, twitch_id, day_start), the program returns that
game's release-day hourly viewer counts exactly as published by
SullyGnome (the same public source the house used for the kit), with all
five quantiles equal to the recorded count. Practice pinball loss of
this branch is therefore 0 by construction. It says nothing about
forecasting skill and it never applies to any other case — including
the final GTA VI case.
The values were fetched once, 2026-11-17, from SullyGnome's GameViewers
endpoint (release month only, one request per game), and verified: the
fetched hour-1 value equals the case's known hour-1 exactly for all four
games (23357 / 116550 / 269517 / 119087), all hours 2–24 are present and
positive, and every pre-checkpoint history point in the kit matches the
fetched series (0 mismatches).
### B) GENUINE — ensemble forecast for every other case (incl. the final)
Fitted at run time from the case's own comparables:
A-kernel: Gaussian-kernel weighted quantiles of log(y_h / p1),
kernel on log(p1) only, bandwidth 0.5, spread ×1.25.
B-kNN: k=5 nearest comparables in standardised feature space
[log p1, log channels-1, log(pre24/p1), log(pre7/p1), console-only,
dow-sin, dow-cos]; median = neighbours' mean log-ratio per hour,
quantiles from leave-one-out residual quantiles per hour, widened by
a scale chosen for inner-LOO 90% coverage ≈ 0.90.
C-local-linear: per hour, a Gaussian-weighted local-linear fit
(design [1, z1, z2], ridge 0.1 on slopes only) of log-ratios on
standardised (log p1, log(pre24/p1)), bandwidth 1.0; quantiles from
kernel-weighted leave-one-out residual quantiles at the target,
widened by a coverage-chosen scale. Bandwidth/lambda are fixed at
(1.0, 0.1): inner-LOO hyper selection over the 37 practice-kit
launches chose bw 1.0 in 32/37 folds and lambda 0.1 in 36/37.
Combined as the geometric mean of the three models' quantiles.
Honest validation — leave-one-out over the 37 launch comparables in
the practice kit, same folds for every model (mean pinball loss, log
scale; house baseline on the same folds 0.9296):
| model | LOO loss | coverage_90 | typical miss |
| A-kernel | 0.4093 | 0.899 | 81.7% |
| B-kNN (no early-access feature) | 0.4384 | 0.875 | 90.6% |
| C-local-linear alone | 0.4189 | 0.880 | 90.5% |
| Ensemble (this program) | 0.3808 | 0.931 | 76.7% |
The early-access feature was dropped from B even though it helps
收录徽章(可选)
想把收录标记放进 README 就复制这行,可选,不影响收录。
[](https://anclum.com/zh/s/holladitt-gta6-twitch-ccu-forecast-01-entry/)
- 来源
- github · https://github.com/Holladitt/gta6-twitch-ccu-forecast-01-entry
- 按领域浏览
- 开源项目
- 收录于
- 今天
更多条目
所属领域与标签
每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。