Daily build 2026-08-28

README.md

@@ -39,8 +39,8 @@ configs:
 # Polymarket Order Book Dataset
 
 Order-book snapshots from a prediction market, collected continuously between
-**2026-07-10** and **2026-08-26**: `341,392,814` quote observations across
-`190,813` markets, plus settlement outcomes and a separate high-frequency feed
+**2026-07-10** and **2026-08-27**: `347,645,386` quote observations across
+`202,590` markets, plus settlement outcomes and a separate high-frequency feed
 that records actual traded prices.
 
 It is published so other people can build and train on it without first spending
@@ -72,16 +72,16 @@ import duckdb
 duckdb.sql("SELECT * FROM 'polymarket-data/quotes/**/*.parquet' LIMIT 5").show()
 ```
 
-**GitHub Releases** (a single dated tarball, ~172 MB):
+**GitHub Releases** (a single dated tarball, ~176 MB):
 
 ```bash
-gh release download data-2026-08-27 --repo DineshKumar8399/polymarket-orderbook-dataset
+gh release download data-2026-08-28 --repo DineshKumar8399/polymarket-orderbook-dataset
 tar --zstd -xf polymarket-orderbook-*.tar.zst
 ```
 
-Each release is a frozen snapshot, so `data-2026-08-27` is reproducible: cite the
+Each release is a frozen snapshot, so `data-2026-08-28` is reproducible: cite the
 tag and anyone can reconstruct the exact data you trained on. Latest build:
-**2026-08-27**.
+**2026-08-28**.
 
 ---
 
@@ -89,13 +89,13 @@ tag and anyone can reconstruct the exact data you trained on. Latest build:
 
 | File | Rows | What it is |
 |---|---|---|
-| `quotes/dt=YYYY-MM-DD/*.parquet` | `341,392,814` | Book quotes for every tracked market, partitioned by date |
-| `markets.parquet` | `190,813` | One row per market: question text, category, coverage |
-| `labels.parquet` | `157,956` | Binary settlement outcomes, with a `source` column |
-| `watch_quotes.parquet` | `2,038,730` | High-frequency feed — **the only table with traded prices** |
+| `quotes/dt=YYYY-MM-DD/*.parquet` | `347,645,386` | Book quotes for every tracked market, partitioned by date |
+| `markets.parquet` | `202,590` | One row per market: question text, category, coverage |
+| `labels.parquet` | `158,380` | Binary settlement outcomes, with a `source` column |
+| `watch_quotes.parquet` | `2,086,258` | High-frequency feed — **the only table with traded prices** |
 | `data_quality.parquet` | 7 | The known issues below, as queryable rows |
 
-Total: about `172 MB` of ZSTD-compressed Parquet.
+Total: about `176 MB` of ZSTD-compressed Parquet.
 
 ---
 
@@ -174,7 +174,7 @@ also the cost per unit of payoff.
 | `first_ts` / `last_ts` | timestamp | Coverage window |
 
 `question` is stored here rather than on every quote row — repeating it
-`341,392,814` times is most of why the raw CSV was 32 GB.
+`347,645,386` times is most of why the raw CSV was 32 GB.
 
 ### `labels.parquet`
 
@@ -252,7 +252,7 @@ has been **dropped rather than shipped as an empty column named `last`**.
 
 If your question is "did this actually transact" — fill realism, execution
 modelling, print-versus-quote — it is only answerable on `watch_quotes`, which
-covers 5,699 markets rather than 190,813. Note `last_traded` is itself 83.7%
+covers 5,826 markets rather than 202,590. Note `last_traded` is itself 83.3%
 populated, not 100%.
 
 ### 4. Crossed books
@@ -264,8 +264,8 @@ your method is sensitive to it.
 ### 5. Labels are time-censored
 
 Most `source = 'api'` labels come from a one-off backfill run on 2026-07-23/24.
-So "has a label" correlates strongly with "settled before Jul 24" — 155,811
-markets (82%) carry an authoritative label, and they are **not a random 82%**.
+So "has a label" correlates strongly with "settled before Jul 24" — 156,225
+markets (77%) carry an authoritative label, and they are **not a random 77%**.
 
 This bites hardest on walk-forward validation: naively splitting train/test on a
 late date can leave you with an empty test set and a script that reports success
@@ -299,13 +299,13 @@ Markets by category:
 
 ```
    category  slugs
-     sports 185728
-   politics   2627
-    climate   1042
-    culture   1008
+     sports 194578
+   politics   5281
+    culture   1243
+    climate   1072
       macro    154
  technology     93
-    finance     82
+    finance     90
      crypto     51
     science     14
 geopolitics     14

labels.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aa768e8e16b68bb8ce8df7ae5189c5a99cc0063c68f719c5faa75732b752fffa
-size 617008
+oid sha256:514294690dd2a65b1cf489062d877860e125e8741c445a6345f18b77e9938c0a
+size 619628

markets.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f6d0c2dddee183d4c1a8ffadc1f227f7c205acd133c9cc18b6dc9dbaf5eaa7cf
-size 2548727
+oid sha256:7dd2a0fa380efeb87f9122ed7e86ff4aa511397554d9461230c825346752aec8
+size 2690538

quotes/dt=2026-08-27/data_0.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:311ed138efc58349b52154f4624978c990906df462816b72e2d6779cd5f85b53
+size 3728840

watch_quotes.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a9ea96d0fecd2f439f1da7c1d31f56891d5da07b9abe227c1e0ff43097d16c05
-size 6382408
+oid sha256:3ff043e3a17212ded4ce0f44e8554f1dc424d90c824dbf0d69ba52c3804a6004
+size 6425723