Datasets:

alexnguyen25
/

patchoracle-skyblock-patch-events

license: cc-by-sa-4.0
language:
  - en
pretty_name: SkyBlock Patch → Price Events
size_categories:
  - 1K<n<10K
tags:
  - event-study
  - game-economy
  - hypixel-skyblock
  - llm-extraction
task_categories:
  - tabular-regression
  - tabular-classification
configs:
  - config_name: default
    data_files:
      - split: train
        path: patchoracle_events.parquet

Dataset card — SkyBlock Patch → Price Events

What: one row per (patch, bazaar item) event: LLM-extracted change facts from Hypixel SkyBlock patch notes (2019–2026) joined to market outcomes (raw and market-adjusted % price moves at D+1/D+3/D+7).

Why it exists: no labeled dataset connects game patch text to in-game market reactions. Prior public work predicts next-tick prices from price history alone (autocorrelation-trivial). This dataset supports event-driven prediction: given what a patch says, what happens to the affected items?

⚠️ v3 correction notice (2026-08-03) — read before using v2 numbers

v3 replaces v2. If you pulled v2, re-pull. v2's row counts were wrong and its central caveat was unquantified. Both are fixed here, and several claims the project made about this data have since been retracted by the project itself.

	v2 (published 2026-08-01)	v3 (this release)
events	2,599	2,078
usable labels (`excluded = false`)	1,403	1,087
`name-match-seed` rows	1,090	569 (405 usable)
rows with a NULL `excerpt`	1,094	0
seed false-positive risk	prose warning only	measured, per-row column

The v2 counts predate the G12 matcher fix. v2's name-match-seed rows contained fabricated attributions the card only warned about in the abstract — most notoriously a Y event mined from the phrase "Howdy y'all!". Those are gone (see name_match_risk below), every row now carries the excerpt it came from, and the residual risk is a column you can filter on rather than a paragraph you have to trust.

Retractions affecting claims made about this dataset. The data is sound; the trading claims built on it were not. Stated plainly because earlier releases of this project asserted otherwise:

The trading-cost assumption was ~20× too optimistic. Published net figures assumed a 1.0% bid-ask spread proxy. The measured median bazaar spread is 19.5% row-level / 23.4% per item. Treat any net-return figure from earlier releases as gross.
No execution regime yields a profitable strategy. Priced on real bid/ask: crossing the book nets −52.9%; resting both legs earns the spread rather than the signal (return correlates r=0.57 with spread width); resting one leg same-side nets −32.8% on the fills you actually get, with −41.3pp of adverse selection. Do not cite this dataset as supporting a profitable strategy.
Direction is a well-powered null. Directional accuracy is ≈0.52–0.54 against a 0.524 always-predict-down constant; a 96-cell, 48-hour event-time scan found no cell above baseline. Magnitude reacts; direction does not. The defensible predictive result is that post-patch price range is forecastable from an item's own 30-day trailing volatility at ρ ≈ 0.45–0.52 out-of-sample — and adding patch-note features does not measurably improve it.
A published fill-rate statistic was a tautology. A "100% fill rate, 0.00pp adverse selection" figure was an algebraic identity — its threshold was a member of the set being minimised over — not a measurement. Retracted; the real entry fill rate is 52.4% with −2.77pp adverse selection.

Size (v3)

2,078 events across 232 patches and 596 bazaar items; 1,087 rows carry usable outcome labels (excluded = false). The source corpus is 396 patches (2019-06-11 → 2026-07-15); patches touching no bazaar-resolvable item contribute no rows.

Fields

Column	Meaning
`patch`, `patch_date`, `source`, `forum_url`	which patch (wiki changelog title or RSS), when, provenance
`anchor_ts`, `anchor_precision`	resolved publish timestamp. `hour` = real forum post time (152 patches); `day` = placeholder `00:00 UTC`, not a measured time (244 patches)
`event_id`	stable row id (matches `event_id` in the project's public API)
`item_tag`	CoflNet/Hypixel bazaar tag (e.g. `ENCHANTED_DIAMOND`)
`change_type`	buff / nerf / new / removed / fix / mention (from text only)
`mechanism`	drop_rate / recipe / shop_price / stats / availability / other
`magnitude_text`, `magnitude_pct`	magnitude as written in the patch; parsed % when stated
`affects_supply`, `affects_demand`	which side of the market the text says is touched
`is_new_content`	item introduced by this patch (no pre-history ⇒ excluded from labels)
`is_new_content_contradicted`	NEW in v3. `true` when `is_new_content` is set but the item has price rows before its own patch date — i.e. the flag is wrong. True for 77 of 339 (23%).
`extraction_model`	which extractor produced the row — see provenance
`excerpt`	the sentence(s) the extraction came from. Non-null for all 2,078 rows in v3
`name_match_risk`	NEW in v3. Per-row attribution audit — see below
`pre_price`, `pre_vol_pct`, `pre_drift_pct`, `pre_volume`	pre-anchor market state (strictly before the patch)
`ret_d{1,3,7}_pct`	raw % move vs D-1 close, winsorized ±50
`abn_d{1,3,7}_pct`	market-adjusted (abnormal) % move vs a volume-weighted bazaar index
`d7_censored`	another patch touched the item inside the D+7 window
`excluded`, `exclude_reason`	rows without valid labels (new items, thin history, volume floor)

`name_match_risk` — v2's caveat, now measured

v2 warned that seed rows "include false-positive mentions" without saying how many. With excerpts stored, that is checkable. Two distinct classes, and only the second still exists:

value	rows (all / usable)	meaning
`not_applicable`	1,509 / 682	LLM-typed row — a model read the passage and asserted a change, so the name-matching bug never applied
`wholeword_ok`	499 / 338	seed row; item name appears as a whole word in the excerpt
`homonym_sense_ambiguous`	70 / 67	seed row whose name is a common English word, so the match may be the wrong sense
`substring_no_wholeword_match`	0 / 0	class eliminated. This is the G12 bug ("Howdy y'all!" → `Y`); no row exhibits it

The homonym_sense_ambiguous class is mixed, not uniformly wrong. Verified by hand:

wrong sense: DIAMOND from "Wooden, Gold, and Diamond chests"; ICE from "Ice Essence Shop"; LOG from "You will need to be MVP+ … Log in"
correct: FLINT from "Cobblestone / Stone … Gravel / Flint"; WHEAT from "Fixed [[Wheat]] growing on…"

They are flagged, not dropped — dropping them would discard real rows and diverge from the project's own labels. For a conservative subset use name_match_risk != 'homonym_sense_ambiguous' (1,020 usable rows); otherwise treat them as weak labels.

Method note, so you can reproduce or distrust it: the check is a whole-word regex on the item's display name, tolerating a plural/possessive on the final word. That tolerance is load-bearing — a first version without it flagged 63 rows as fabricated that were merely plurals ("Iron Ingots", "Bejeweled Handles", "red gifts"), i.e. it reported a ~6% fabrication rate that was entirely its own bug. Code: data_release/export.py::classify_name_match.

Construction & caveats

Patch text: Hypixel SkyBlock Wiki Changelog/* pages (CC-BY-SA) via the MediaWiki API; live patches from the official forum RSS.
Prices: CoflNet SkyApi historical bazaar series (2h/daily buckets). This release contains only per-event derived aggregates, not CoflNet's raw series — pull those from their API directly.
Extraction: Claude Sonnet 5 (claude:claude-sonnet-5), JSON-schema constrained facts-only prompt (no outcome-flavored fields, to avoid pretraining-knowledge leakage), patch text chunked at 15k chars; item names resolved against the 1,609-tag bazaar vocabulary.
Row provenance (extraction_model): claude:claude-sonnet-5 1,350 — typed events; name-match-seed 569 — item name matched in patch text but no extractor typed a change for it; ollama:qwen3 159 — v1 events the v2 pass did not re-find, retained rather than deleted.
Timestamps are a known weak point. 244 of 396 patches sit at anchor_precision = 'day', anchored at 00:00 UTC — a fiction; real patches publish at a median ~15:00 UTC. Always split intraday results by anchor_precision: pooling the two hid a genuine effect in this project's own analysis. ~86.7% of the missing timestamps are recoverable (the wiki stores the forum thread as an {{Update Link|post=}} template, which the scraper's literal-URL regex could not see) but that is not applied in v3.
Exclusions are selective, not random. 991 of 2,078 events are excluded: 485 thin pre-history, 339 new content, 167 below volume floor. The excluded new-content slice is 1.5–2.9× more dispersed in outcome magnitude than a comparable non-event window, and that survives matching on liquidity. So the usable subset is systematically tamer than the full event population — material if you model magnitude, immaterial for direction (null anyway).
The market index is not a tradeable benchmark. abn_* columns are adjusted against a volume-weighted bazaar index that compounds at ~+19,658%/yr with a median breadth of 21 items/day. Use abn_* for event-study framing and ret_* for anything resembling P&L.
Other noise sources (quantified in the repo's research/02): market-wide moves (mitigated by abnormal returns), Alpha-server anticipation (patches are partially priced in before release), thin-book manipulation (volume floor + winsorization), overlapping patches (d7_censored).

Extraction-quality findings (v2 re-extraction, retained)

Two pre-committed evaluations; they point in different directions and both are disclosed.

Extraction quality: improved. A blinded LLM-judge pass over a 44-patch stratified sample scored Sonnet 5 well above qwen3 on both axes — correctness 4.30 vs 2.73, completeness 2.61 vs 1.36 (0–5 scale). Pairwise change_type agreement is 0.75–0.85 between independent Claude tiers vs 0.45–0.51 between qwen3 and any Claude tier.
Downstream model gate: marginal FAIL. Retraining on v2 data moved test directional accuracy 0.596 → 0.573 (−2.3pt, past a pre-committed −2pt gate) while three other tracked metrics held or improved. Don't cite this dataset as having improved downstream accuracy — the measured claim is better extraction, not better predictions.
Class-distribution shift. Claude labels 45% of typed events buff (qwen3: 23%) and produces almost no removed events (1 vs 83). v1 → v2 is a real distribution break if you use change_type frequencies as features or priors.

Companion datasets

patchoracle-skyblock-mayor-cells — the 42 pre-registered (basket × horizon) cells from the mayor-election event study, with verdicts and permutation p-values, failures included.
patchoracle-skyblock-patches — the 396-patch corpus metadata with resolved anchors, useful independently of the event labels.

License & attribution

Labels/extractions: CC-BY-SA 4.0 (inherits the wiki's license). Please credit: Hypixel SkyBlock Wiki (Fandom), CoflNet SkyApi, Hypixel API. Not affiliated with Hypixel. Built by the PatchOracle project.

Transparency notes

Anchor correction: an early pipeline version could anchor at patch-day close when D-1 was missing (a post-patch leak); fixed and fully relabeled. This release is post-fix.
Mega-patch gap (v1): CLOSED. The one patch that resisted local-LLM extraction in v1 (Changelog/2024/April 19, 46k chars) extracts cleanly under Claude. All corpus patches processed with zero failures.
Every negative result above is published because it was measured, not because it was convenient. If a claim in this card is not supported by the data, that is a bug — please open a discussion.