Blame for data_processing.md

# Data Processing Documentation

**Dataset**: StakeBench — Evaluating Language Understanding Grounded in Market Commitment (working name during collection: PolySent)
**Updated**: 2026-10-27 (v1.1: §6 now points to the released raw files and scripts; collection documentation unchanged since 2026-03-18)
**Platforms**: Polymarket (`polymarket_data/`) · Manifold Markets (`manifold_data/`)

---

## Table of Contents

1. [Overview](#1-overview)
2. [Polymarket Dataset](#2-polymarket-dataset)
   - 2.1 [Topic Collection (`get_polymarket_topics.py`)](#21-topic-collection)
   - 2.2 [Comment Collection (`get_polymarket_comments.py`)](#22-comment-collection)
   - 2.3 [Cleaning (`polymarket_data/clean_comments.py`)](#23-cleaning)
3. [Manifold Dataset](#3-manifold-dataset)
   - 3.1 [Topic Collection (`get_manifold_topics.py`)](#31-topic-collection)
   - 3.2 [Comment Collection (`get_manifold_comments.py`)](#32-comment-collection)
   - 3.3 [Cleaning (`manifold_data/clean_comments.py`)](#33-cleaning)
4. [Schema Comparison](#4-schema-comparison)
5. [Descriptive Statistics](#5-descriptive-statistics)
   - 5.1 [Basic Scale](#51-basic-scale)
   - 5.2 [Market Coverage](#52-market-coverage)
   - 5.3 [Comment Text Analysis](#53-comment-text-analysis)
   - 5.4 [Position (Stake) Data](#54-position-stake-data)
   - 5.5 [Reply Structure](#55-reply-structure)
   - 5.6 [Engagement (Reactions/Likes)](#56-engagement-reactionsikes)
   - 5.7 [User Behavior](#57-user-behavior)
   - 5.8 [Temporal Distribution](#58-temporal-distribution)
   - 5.9 [Category Distribution](#59-category-distribution)
   - 5.10 [Manifold-Specific: Resolution & Probability](#510-manifold-specific-resolution--probability)
   - 5.11 [Cross-Platform Comparison](#511-cross-platform-comparison)
6. [Raw Files and Scripts](#6-raw-files-and-scripts)

---

## 1. Overview

| | Polymarket | Manifold |
|---|---|---|
| Market type | Real-money binary/multi-outcome prediction markets | Virtual-currency (Mana) binary prediction markets |
| API | `gamma-api.polymarket.com` + `data-api.polymarket.com` | `api.manifold.markets/v0` |
| Market filter (collection) | Event-level: total volume ≥ $10,000 USD + comments ≥ 100 | Market-level: volume ≥ 10,000 M$ + bettors ≥ 50 + has comments |
| Comment filter (collection) | ≥ 3 chars; ≥ 10 per market | ≥ 3 chars; ≥ 10 per market; within market lifecycle |
| Comment filter (cleaning) | Markets with < 10 comments removed; empty questions removed | Markets with < **100** comments removed; empty questions removed |
| **Cleaned comments** | **514,314** | **46,562** |
| **Cleaned markets** | **2,070** | **191** |
| **Unique users** | **24,909** | **3,250** |
| Date range | 2024-05-02 → 2026-03-12 | 2022-01-14 → 2026-03-07 |

---

## 2. Polymarket Dataset

### 2.1 Topic Collection

**Script**: `get_polymarket_topics.py`
**Output**: `polysent_data/valid_topics.jsonl`, `polysent_data/valid_topics_summary.csv`

#### Pipeline

```
GET gamma-api.polymarket.com/markets
  └─ params: closed=True, limit=500, offset=0..N
  └─ paginate until empty response (≤ MAX_PAGES=2000, ~1M markets)
  └─ per market: extract event_id, volume, comment_count
  └─ aggregate by event_id:
       total_volume  = Σ volume across all markets in the event
       comment_count = event-level commentCount (shared)
       best_market   = highest-volume market (representative)
  └─ filter: total_volume ≥ $10,000  AND  comment_count ≥ 100
  └─ classify category by keyword matching (crypto/politics/economics/sports/technology)
```

#### Key Parameters

| Parameter | Value | Description |
|---|---|---|
| `MIN_TOTAL_VOLUME` | 10,000 | Minimum USD trading volume per event |
| `MIN_COMMENT_COUNT` | 100 | Minimum comments per event |
| `PAGE_LIMIT` | 500 | Markets per API request |
| `SLEEP` | 0.15 s | Inter-request delay |

#### Category Classification

Keywords are matched against the representative market's question text (case-insensitive substring match). The category with the most keyword hits wins; no match → `"other"`.

Categories: `crypto`, `politics`, `economics`, `sports`, `technology`, `other`

---

### 2.2 Comment Collection

**Script**: `get_polymarket_comments.py`
**Input**: `polysent_data/valid_topics_summary.csv` (column `best_market_id`)
**Output**: `polysent_data/comments.csv`

#### Pipeline

```
For each target market_id:
  1. GET gamma-api.polymarket.com/comments
       └─ params: parent_entity_type=market, parent_entity_id=<id>
       └─ fallback: Event → Series (auto-detected)
       └─ paginate with offset until exhausted

  2. For each commenter's user_address:
       GET data-api.polymarket.com/activity
         └─ parallel fetch (8 threads, ThreadPoolExecutor)
         └─ reconstruct historical position at comment creation time:
              position_side  = Yes / No / <team> (market-specific)
              position_size  = USD net cost (Σ BUY − Σ SELL; can be negative when profitable)
              position_shares = net shares held

  3. Write rows to comments.csv (append / resumable via progress.json)
```

#### Anti-Ban Measures

- Random User-Agent rotation (5 browser strings)
- Exponential backoff + jitter on HTTP 429/5xx (max 120 s)
- Optional proxy pool rotation (`proxies.txt`)
- Inter-request jitter sleep: `BASE_SLEEP ± JITTER = 0.12 ± 0.08 s`
- Resume support: `progress.json` tracks `done / skipped / failed` per market

#### Collection Filters

| Filter | Value | Note |
|---|---|---|
| `MIN_COMMENT_LENGTH` | 3 chars | Loose; stricter post-hoc filter recommended (>20) |
| `MIN_COMMENTS_PER_MARKET` | 10 | Markets below this are marked `skipped` |
| Time window | `[market_start_time, market_end_time]` | Comments outside market lifecycle dropped |

#### Output Schema (raw `comments.csv`)

| Field | Type | Description |
|---|---|---|
| `market_id` | str | Polymarket market ID |
| `slug` | str | URL slug |
| `question` | str | Market question |
| `condition_id` | str | On-chain condition ID |
| `category_validated` | str | Keyword-classified category (5 classes) |
| `market_start_time` | ISO8601 | Market open time |
| `market_end_time` | ISO8601 | Market close/settlement time |
| `parent_entity_type_used` | str | `Event` or `Series` (fallback tracking) |
| `parent_entity_id_used` | str | Entity ID used for comment API call |
| `comment_id` | str | Unique comment ID (may repeat across markets sharing a Series) |
| `created_at` | ISO8601 | Comment timestamp (UTC) |
| `text` | str | Comment body |
| `pseudonym` | str | Anonymized display name |
| `user_address` | str | On-chain ETH wallet address |
| `proxy_wallet` | str | Polymarket proxy wallet |
| `position_side` | str | Yes / No / team name at comment time |
| `position_size` | float | USD net cost (can be negative when profitable) |
| `position_shares` | float | Net shares held at comment time |
| `reactions` | JSON str | Raw reaction array |
| `reaction_count` | int | Total reactions (occasional −1 from API bug) |
| `report_count` | int | Number of reports |
| `parent_comment_id` | str | Parent comment ID (NaN for top-level) |
| `reply_count` | int | Replies received (within collection window) |

> **Note on `comment_id` duplicates**: Markets sharing a Polymarket Series pool share the same comment thread. The script collects comments for each market with its own time window, so the same `comment_id` may appear with different `market_id` values. Deduplicate by `comment_id` for unique-comment analysis.

---

### 2.3 Cleaning

**Script**: `polymarket_data/clean_comments.py`

| Step | Operation | Effect |
|---|---|---|
| 1 | Drop `category` column | 100% missing (Gamma API does not expose this field) |
| 2 | Strip whitespace from `question`; remove markets where question is empty/`"nan"` | Data integrity |
| 3 | Remove markets with < 10 valid comments | Minimum discussion threshold |
| 4 | Reorder columns to reference schema order | Consistency |

**Run summary (from `run.log`)**:
- Input: 2,469 markets from `valid_topics_summary.csv`
- Completed: 2,070 / Skipped: 399 / Failed: 0
- Runtime: ~9.7 hours

---

## 3. Manifold Dataset

### 3.1 Topic Collection

**Script**: `get_manifold_topics.py`
**Output**: `manifold_data/valid_markets.jsonl`, `manifold_data/valid_markets_summary.csv`

#### Pipeline

```
GET api.manifold.markets/v0/search-markets
  └─ params: filter=resolved, contractType=BINARY, sort=liquidity
  └─ paginate with offset (100 per page, up to MAX_PAGES=5000)
  └─ per market apply filters:
       volume   ≥ 10,000 M$ (Mana)
       bettors  ≥ 50
       lastCommentTime is not None
  └─ classify category by keyword matching (same keyword dict as Polymarket)
  └─ NOTE: comment-count filter deferred to cleaning step
```

#### Key Parameters

| Parameter | Value | Description |
|---|---|---|
| `MIN_VOLUME_MANA` | 10,000 | Minimum Mana (M$) trading volume |
| `MIN_BETTORS` | 50 | Minimum unique bettors |
| `PAGE_LIMIT` | 100 | Markets per API request |
| `BASE_SLEEP` | 0.15 s | Inter-request delay (Manifold limit: 500 req/min) |

> **Currency note**: Manifold uses virtual currency Mana (M$), not real USD. `volume` and `position_size` are not directly comparable to Polymarket figures.

---

### 3.2 Comment Collection

**Script**: `get_manifold_comments.py`
**Input**: `manifold_data/valid_markets.jsonl`
**Output**: `manifold_data/comments.csv`

#### Pipeline

```
For each market (resolution = YES or NO only):
  1. GET api.manifold.markets/v0/comments
       └─ params: contractId=<id>, limit=1000
       └─ cursor-based pagination (before=<last_id>)
       └─ extract plain text from ProseMirror JSON (recursive content node walk)
       └─ filter: created_at ∈ [market.created_time, market.resolution_time]

  2. For each unique commenter's userId (parallel, 8 threads):
       GET /v0/bets?contractId=<id>&userId=<uid>
         └─ cursor-based pagination
         └─ reconstruct position at comment time:
              net_mana[YES] = Σ amount for YES bets before comment_ts
              net_mana[NO]  = Σ amount for NO bets before comment_ts
              position_side = dominant side (or None if both ≤ 0)
         └─ prob_at_comment = probAfter of last bet before comment_ts

  3. Append rows to comments.csv (resumable via progress.json)
```

#### Key Differences from Polymarket

| Aspect | Polymarket | Manifold |
|---|---|---|
| Position currency | USD (real money) | Mana (virtual currency) |
| Position source | Dedicated activity API per user | Bet records reconstructed per user |
| Market probability | Dedicated `/prices-history` endpoint | Interpolated from `probAfter` in bet records |
| Text format | Plain text | ProseMirror JSON → recursively extracted |
| Reply count | Provided by API | Not in API; back-calculated from `replyToCommentId` |
| Resolution ground truth | Not in API | `YES` / `NO` in market metadata |

#### Collection Filters

| Filter | Value |
|---|---|
| `MIN_COMMENT_LENGTH` | 3 chars |
| `MIN_COMMENTS_PER_MARKET` | 10 |
| Resolution filter | Only `YES` / `NO` markets (excludes `MKT`, `CANCEL`, `N/A`) |
| Lifecycle filter | Comments outside `[created_time, resolution_time]` dropped |

#### Output Schema (raw `comments.csv`)

| Field | Type | Description |
|---|---|---|
| `market_id` | str | Manifold market ID |
| `slug` | str | URL slug |
| `question` | str | Market question |
| `outcome_type` | str | Always `BINARY` |
| `category` | str | Keyword-classified category |
| `resolution` | str | `YES` / `NO` (ground truth) |
| `comment_id` | str | Unique comment ID |
| `created_at` | ISO8601 | Comment timestamp (UTC, ms precision) |
| `text` | str | Extracted plain text |
| `user_id` | str | Manifold user ID |
| `user_name` | str | Username |
| `position_side` | str | `YES` / `NO` / empty |
| `position_mana` | float | Net Mana invested at comment time |
| `position_shares` | float | Net shares held at comment time |
| `prob_at_comment` | float | Market probability snapshot at comment time |
| `likes` | int | Like count |
| `parent_comment_id` | str | Parent comment ID (NaN for top-level) |
| `reply_count` | str | Empty (back-calculated during cleaning) |

---

### 3.3 Cleaning

**Script**: `manifold_data/clean_comments.py`

| Step | Operation | Effect |
|---|---|---|
| 1 | Back-calculate `reply_count` from `parent_comment_id` value counts | API does not provide `reply_count` |
| 2 | Rename columns to align with Polymarket schema | `user_id→user_address`, `user_name→pseudonym`, `position_mana→position_size`, `likes→reaction_count` |
| 3 | Drop markets where `question` is empty/`"nan"` | Data integrity |
| 4 | **Remove markets with < 100 valid comments** | Ensures sufficient discussion depth per market for NLP tasks |
| 5 | Reorder columns (retain Manifold-specific fields: `outcome_type`, `resolution`, `prob_at_comment`) | Schema alignment |

> **Threshold change**: The minimum comments per market was raised from 10 to **100** to retain only well-discussed markets with sufficient context for sentiment analysis. This reduces market count from ~2,926 to **191** but increases average comment depth from 21 to **163** per market.

---

## 4. Schema Comparison

| Field | Polymarket | Manifold | Notes |
|---|---|---|---|
| `market_id` | ✓ | ✓ | Different ID systems (numeric vs. alphanumeric) |
| `slug` | ✓ | ✓ | |
| `question` | ✓ | ✓ | |
| `condition_id` | ✓ | — | On-chain condition (Polymarket only) |
| `outcome_type` | — | ✓ | Always `BINARY` (Manifold only) |
| `category_validated` / `category` | ✓ | ✓ | Same keyword logic; different field names |
| `resolution` | — | ✓ | `YES`/`NO` ground truth (Manifold only) |
| `market_start_time` | ✓ | — | |
| `market_end_time` | ✓ | — | |
| `parent_entity_type_used` | ✓ | — | API fallback tracking (Polymarket only) |
| `comment_id` | ✓ (may repeat) | ✓ (unique) | |
| `created_at` | ✓ | ✓ | ISO8601 UTC |
| `text` | ✓ | ✓ | |
| `pseudonym` | ✓ | ✓ | (renamed from `user_name`) |
| `user_address` | ✓ | ✓ | (renamed from `user_id`) |
| `proxy_wallet` | ✓ | — | Polymarket on-chain proxy |
| `position_side` | Yes/No/team | YES/NO | Format differs; Polymarket has 73 distinct values |
| `position_size` | USD (can be negative) | Mana M$ | Not directly comparable |
| `position_shares` | ✓ | ✓ | |
| `prob_at_comment` | — | ✓ | Market probability snapshot (Manifold only) |
| `reactions` | JSON str | — | Raw reactions (Polymarket only) |
| `reaction_count` | ✓ | ✓ | (renamed from `likes`); Polymarket occasionally −1 |
| `report_count` | ✓ | — | |
| `parent_comment_id` | ✓ | ✓ | |
| `reply_count` | ✓ | ✓ | Back-calculated for Manifold |

---

## 5. Descriptive Statistics

> Analysis scripts: `data/analysis/descriptive_stats.py` and `data/analysis/plot_stats.py`
> Run from `data/` directory: `python analysis/descriptive_stats.py`

---

### 5.1 Basic Scale

| Metric | Polymarket | Manifold |
|---|---|---|
| Total comments (cleaned) | **514,314** | **46,562** |
| Unique `comment_id` (deduplicated) | **266,446** | 46,562 |
| Unique markets | **2,070** | **191** |
| Unique users | **24,909** | **3,250** |
| Fields per row | 23 | 18 |
| Date range | 2024-05-02 → 2026-03-12 | 2022-01-14 → 2026-03-07 |
| Time span | ~22 months | ~50 months |

> **Polymarket duplicate rows**: 48.2% of rows share a `comment_id` with another row (same comment collected under different `market_id` for markets sharing a Series comment pool). Deduplicate with `df.drop_duplicates(subset='comment_id', keep='first')` before unique-comment analyses.

---

### 5.2 Market Coverage

#### Comments per Market

| Statistic | Polymarket | Manifold |
|---|---|---|
| Count (markets) | 2,070 | 191 |
| Mean | 248.5 | 243.8 |
| Std | 1,284.1 | 197.1 |
| Min | 10 | 100 |
| 25th pct | 41 | 126 |
| Median | 70 | 163 |
| 75th pct | 127 | 275 |
| Max | 25,049 | 990 |

#### Market Depth Brackets (Polymarket)

| Threshold | Markets | Share |
|---|---|---|
| ≥ 10 comments | 2,070 | 100% |
| ≥ 50 comments | 1,363 | 65.8% |
| ≥ 100 comments | 692 | 33.4% |
| ≥ 200 comments | 259 | 12.5% |

#### Top 10 Markets by Comment Count

**Polymarket:**

| Rank | Market ID | Comments | Question |
|---|---|---|---|
| 1 | 519075 | 25,049 | Will Hunor Kelemen win the Romanian presidential election? |
| 2 | 516767 | 22,444 | Will Pierre Poilievre be the next Canadian Prime Minister? |
| 3 | 514139 | 19,284 | Will Rafał Trzaskowski be the next President of Poland? |
| 4 | 253591 | 17,034 | Will Donald Trump win the 2024 US Presidential Election? |
| 5 | 546814 | 16,326 | Will Zelenskyy wear a suit before July? |
| 6 | 506968 | 14,810 | Will Israel invade Syria in 2024? |
| 7 | 507276 | 12,949 | TikTok banned in the US before May 2025? |
| 8 | 532741 | 11,074 | US military action against Iran before July? |
| 9 | 551650 | 11,037 | US military action against Iran before August? |
| 10 | 552650 | 9,331 | Fordow nuclear facility destroyed before July? |

**Manifold:**

| Rank | Market ID | Comments | Question |
|---|---|---|---|
| 1 | l2Pci4rwD2S50eFGZiCj | 990 | Is the probability of dying in the Snake Eyes Paradox 1/36? |
| 2 | 1sfbjcTu6zgqjCVjqWA2 | 961 | Will LK-99 room-temperature superconductivity be confirmed? |
| 3 | 7RZ7CnNVPKOGSh3v5pnn | 933 | Will @firstuserhere walk 100,000 steps in a day before 2023? |
| 4 | d1t4k3nz3t | 927 | Will Trump win the 2024 Election? |
| 5 | gPA029Rcyh | 919 | Will Manifold solve my puzzle? |
| 6 | YTIuuSsNRn2OlA4KykRM | 918 | Will Biden be the 2024 Democratic Nominee? |
| 7 | U3zLgOZkGUE7cvG98961 | 891 | Whales vs. Minnows: Will traders hold at least 10000x as many shares? |
| 8 | 0EVOQyEBz624ooYpOCfc | 844 | Hamas command center under al-Shifa Hospital? |
| 9 | PXd5l5qBVArJrSYZhivM | 808 | Will Tumbles be late to pay back a loan? |
| 10 | NRLwgQYqCLk5zTasmxOn | 762 | Will the average percentage of this market be above 50%? |

#### Polymarket: parent_entity_type Distribution

| Type | Comments | Share |
|---|---|---|
| Series | 323,082 | 62.8% |
| Event | 191,232 | 37.2% |

Series markets (comment pool shared across multiple related markets) account for 62.8% of all rows and are the primary source of `comment_id` duplication.

---

### 5.3 Comment Text Analysis

#### Text Length (characters)

| Statistic | Polymarket | Manifold |
|---|---|---|
| Mean | 69.4 | 208.5 |
| Std | 87.9 | 347.5 |
| Min | 2 | 3 |
| 25th pct | 22 | 49 |
| Median | 44 | 107 |
| 75th pct | 85 | 230 |
| 90th pct | ~160 | ~470 |
| Max | 1,492 | 9,132 |

Manifold comments are ~2.4× longer on average (median 107 vs. 44 chars). This reflects Manifold's community culture of extended argumentative discussion, compared to Polymarket's shorter reaction-style comments.

#### Text Length Brackets

| Bracket | Polymarket | Manifold |
|---|---|---|
| < 20 chars | ~124,000 (24.1%) | ~5,600 (12.0%) |
| 20–100 chars | ~243,000 (47.3%) | ~14,900 (32.0%) |
| > 100 chars | ~147,000 (28.6%) | ~26,100 (56.1%) |

> Polymarket `MIN_COMMENT_LENGTH=3` (not the Guide-recommended 20). For downstream tasks requiring substantive text, apply `len(text) > 20` post-hoc filter.

---

### 5.4 Position (Stake) Data

Position data links each comment to the commenter's financial stake at the time of writing — a key feature for the Polymarket Sentiment Analysis (PSA) task.

#### Coverage Summary

| Metric | Polymarket | Manifold |
|---|---|---|
| Comments with `position_side` | 155,205 (30.2%) | 37,497 (80.5%) |
| Comments with `position_size > 0` | 139,848 (27.2%) | 37,430 (80.4%) |
| Comments with `position_size < 0` | 15,336 (3.0%) | 0 |
| No position (NaN) | 359,109 (69.8%) | 9,065 (19.5%) |

**Manifold's much higher position coverage (80.5%)** reflects the fact that Manifold users generally bet before commenting, whereas many Polymarket users comment without holding a position. The `>= 100` market filter further concentrates well-engaged markets.

#### Position Side Distribution

| Value | Polymarket | Polymarket % | Manifold | Manifold % |
|---|---|---|---|---|
| Yes / YES | 77,062 | 15.0% | 19,258 | 41.4% |
| No / NO | 63,760 | 12.4% | 18,239 | 39.2% |
| Team / other names | ~14,383 | 2.8% | — | — |
| No position (NaN) | 359,109 | 69.8% | 9,065 | 19.5% |

Polymarket `position_side` has 73 distinct values: Yes/No (90.8% of positioned comments), Up/Down (crypto markets), and team names (NBA/MLB sports markets).

#### Position Size Distribution (comments with size > 0)

| Statistic | Polymarket (USD) | Manifold (Mana M$) |
|---|---|---|
| Count | 139,848 | 37,430 |
| Mean | $3,164 | M$24,408 |
| Std | $22,472 | M$182,496 |
| 25th pct | $45.7 | M$123 |
| Median | $242 | M$949 |
| 75th pct | $1,174 | M$6,000 |
| Max | $2,956,218 | M$4,404,784 |

> **Polymarket `position_size < 0` (15,336 rows)**: Legitimate data — occurs when a user's total SELL proceeds exceed BUY costs (i.e., trading profit; the user remains in a net-long position as `position_shares > 0`). Do not remove; add `is_profit_state` flag for downstream handling.

#### Position Size Tier Breakdown (Polymarket, > 0 only)

| Tier | Range (USD) | Count | % of positioned |
|---|---|---|---|
| Retail | < $100 | 52,259 | 37.4% |
| Small | $100–$500 | 35,042 | 25.1% |
| Medium | $500–$1,000 | 14,162 | 10.1% |
| Large | $1,000–$5,000 | 23,865 | 17.1% |
| XLarge | $5,000–$10,000 | 6,315 | 4.5% |
| Whale | $10,000–$50,000 | 6,732 | 4.8% |
| Mega | > $50,000 | 1,473 | 1.1% |

#### Position Size Tier Breakdown (Manifold, > 0 only)

| Tier | Range (M$) | Count | % of positioned |
|---|---|---|---|
| Retail | < M$100 | 8,621 | 23.0% |
| Small | M$100–500 | 7,191 | 19.2% |
| Medium | M$500–1,000 | 3,480 | 9.3% |
| Large | M$1,000–5,000 | 7,981 | 21.3% |
| XLarge | M$5,000–10,000 | 3,008 | 8.0% |
| Whale | M$10,000–50,000 | 4,944 | 13.2% |
| Mega | > M$50,000 | 2,205 | 5.9% |

---

### 5.5 Reply Structure

| Metric | Polymarket | Manifold |
|---|---|---|
| Is a reply (has `parent_comment_id`) | 259,226 (50.4%) | 33,020 (70.9%) |
| Has received replies (`reply_count > 0`) | 101,081 (19.7%) | 7,217 (15.5%) |
| `reply_count` mean | 0.46 | 0.55 |
| `reply_count` max | 152 | 78 |

> Manifold's very high reply rate (70.9%) reflects its forum-style discussion culture. Note that `reply_count` is locally computed (within the collection window), not the API's authoritative total.

---

### 5.6 Engagement (Reactions/Likes)

| Metric | Polymarket (`reaction_count`) | Manifold (`reaction_count` / likes) |
|---|---|---|
| Mean | 1.06 | 1.72 |
| Median | 0 | 1 |
| 75th pct | 1 | 2 |
| Max | 160 | 81 |
| With ≥ 1 reaction | 230,329 (44.8%) | — |
| With ≥ 5 reactions | 26,329 (5.1%) | — |
| `reaction_count = −1` (API anomaly) | 42 | 0 |

> Polymarket `reaction_count = −1` (42 rows) is an API-side bug; replace with 0 during downstream cleaning.

---

### 5.7 User Behavior

#### Activity Distribution

| Statistic | Polymarket | Manifold |
|---|---|---|
| Unique users | 24,909 | 3,250 |
| Mean comments/user | 20.6 | 14.3 |
| Median comments/user | 3 | 2 |
| 75th pct | 10 | 9 |
| Max | 6,054 | 1,194 |

Both platforms show a long-tail distribution. The most active Polymarket user posted 6,054 comments (vs. 279,347 in the earlier v0312 dataset), indicating the v2 collection script substantially reduced bot-like behavior.

#### Top 5 Most Active Users

**Polymarket** (by ETH address prefix):

| Rank | user_address | Comments |
|---|---|---|
| 1 | 0x6e26a51e... | 6,054 |
| 2 | 0x5cc591f9... | 5,844 |
| 3 | 0xf3a34b19... | 4,748 |
| 4 | 0x238b427f... | 4,664 |
| 5 | 0x4711c314... | 4,473 |

**Manifold** (by user ID prefix):

| Rank | user_address | Comments |
|---|---|---|
| 1 | 4JuXgDx47x... | 1,194 |
| 2 | y1hb6k7txd... | 762 |
| 3 | zgCIqq8AmR... | 757 |
| 4 | sCOvPc5J2s... | 698 |
| 5 | uIA81X2m3a... | 653 |

---

### 5.8 Temporal Distribution

#### Monthly Comment Volume

**Polymarket** (2024-05 to 2026-03):

| Month | Comments | Note |
|---|---|---|
| 2024-05 | 9 | Platform cold-start |
| 2024-08 | 372 | |
| 2024-10 | 9,485 | Pre-election ramp-up |
| 2024-11 | 25,143 | US election |
| 2024-12 | 48,980 | Post-election activity |
| 2025-01 | 46,020 | |
| 2025-04 | 43,632 | |
| **2025-06** | **123,040** | **Peak** (Iran/nuclear crisis markets) |
| 2025-07 | 53,050 | |
| 2025-08 | 15,425 | |
| 2025-11 | 407 | Sharp drop in market activity |
| 2025-12 | 6,739 | |
| 2026-01–03 | ~224 | Tail (incomplete months) |

**Manifold** (2022-01 to 2026-03):

| Period | Comments | Note |
|---|---|---|
| 2022-01–10 | < 500/month | Early platform growth |
| 2022-11 | 1,165 | Growth inflection |
| 2023-01–12 | 1,350–3,089/month | Steady growth phase |
| **2023-12** | **3,089** | **Peak year for 2023** |
| 2024-07 | 2,603 | Election season |
| 2024-11 | 1,456 | US election |
| 2025-05 | 1,575 | |
| 2026-01–03 | 654–102 | Recent (within-collection) |

---

### 5.9 Category Distribution

#### Polymarket

| Category | Comments | Share (of all) | Share (deduplicated) |
|---|---|---|---|
| politics | 99,858 | 19.4% | **36.3%** |
| crypto | 50,508 | 9.8% | 9.0% |
| economics | 13,486 | 2.6% | 2.8% |
| sports | 5,500 | 1.1% | 0.4% |
| NaN / other | 345,062 | 67.1% | 51.6% |

> `category_validated` is keyword-matched; 67.1% of rows (51.6% deduplicated) fall outside the five defined categories. The high NaN rate reflects question phrasings that don't match any keyword.

#### Manifold

| Category | Comments | Share |
|---|---|---|
| other | 30,875 | 66.3% |
| politics | 6,252 | 13.4% |
| technology | 3,759 | 8.1% |
| crypto | 3,758 | 8.1% |
| sports | 1,174 | 2.5% |
| economics | 744 | 1.6% |

Manifold `category` is assigned at collection time; `other` dominates because many Manifold questions are platform-community or science/culture topics not covered by the five keyword sets.

---

### 5.10 Manifold-Specific: Resolution & Probability

#### Market Resolution

| Resolution | Comments | Share |
|---|---|---|
| YES | 25,009 | 53.7% |
| NO | 21,553 | 46.3% |

A near-balanced YES/NO split avoids class imbalance issues for binary prediction tasks.

#### Market Probability at Comment Time (`prob_at_comment`)

| Metric | Value |
|---|---|
| Coverage | 40,914 / 46,562 (87.9%) |
| Mean | 0.506 |
| Median | ~0.5 |
| Std | — |

The mean probability of ~0.5 indicates comments are distributed across the full lifecycle of markets, not concentrated near resolution (when probability collapses to 0 or 1).

---

### 5.11 Cross-Platform Comparison

| Metric | Polymarket | Manifold |
|---|---|---|
| Total comments | 514,314 | 46,562 |
| Unique markets | 2,070 | 191 |
| Unique users | 24,909 | 3,250 |
| Median text length | 44 chars | 107 chars |
| Median comments/market | 70 | 163 |
| Position coverage | 30.2% | 80.5% |
| Reply rate (has parent) | 50.4% | 70.9% |
| Reaction mean | 1.06 | 1.72 |
| Ground-truth resolution | ✗ | ✓ |
| Market probability | ✗ | ✓ (87.9% coverage) |
| Currency | Real USD | Virtual Mana |

**Key trade-offs**:
- **Polymarket**: larger scale (11× more comments, 11× more markets), real financial stakes, but lower position coverage and no built-in resolution ground truth.
- **Manifold**: smaller but higher-depth markets (min 100 comments), near-complete position data (80.5%), resolution ground truth, and market probability snapshots; however, Mana is virtual currency, limiting financial signal interpretation.

---

## 6. Raw Files and Scripts

The raw outputs named in §2–§3 (`valid_topics.jsonl`, `valid_markets.jsonl`, `comments.csv`,
`user_positions.csv`) are released unmodified in the `raw/` folder of this dataset; `raw/README.md`
gives their row counts, columns and the exact steps that produce the `*_clean.csv` files. The
collection scripts (`get_*_topics.py`, `get_*_comments.py`, `get_*_position.py`, `get_*_data.py`)
and the cleaning scripts (`clean_comments.py`, `clean_positions.py`) are in the code repository
under `collection/`: https://github.com/PEIYUNHUA/stakebench_code_NeurIPS26.

Note on the released `polymarket/comments_clean.csv`: it is GB18030-encoded and 32 rows shorter
than the raw crawl output; the raw file is UTF-8. See `raw/README.md` (*Encoding note*).

---

*Generated: 2026-03-18*