mlsys-io/EventXBench

EventX

A multimodal benchmark linking 9M Twitter/X posts to 11,952 Polymarket prediction markets (2021--2026).

Dataset Description

EventX connects social media posts on Twitter/X to prediction market dynamics on Polymarket. It provides seven tasks spanning two tiers:

• Resolution tier (human-annotated): Post-to-Market Linking (T2), Evidence Grading (T3)
• Forecast tier (deterministic labels): Market Volume Prediction (T1), Market Movement Prediction (T4), Impact Persistence (T5), Cross-Market Propagation (T6)

Supported Tasks

Config	Task	Description
t1	Market Volume Prediction	Predict eventual trading volume from pre-market tweets
t2	Post-to-Market Linking	Match a tweet to the correct prediction market
t3	Evidence Grading	Grade tweet relevance to a market (0-5)
t4	Market Movement Prediction	Predict price direction and magnitude at 2h horizon
t5	Volume & Price Impact	Predict price_impact and volume_multiplier (continuous)
t6	Cross-Market Propagation	Predict spillover to sibling markets
t7	Impact Persistence (Decay)	Classify decay: transient/sustained/reversal
posts	Tweet Metadata	Tweet IDs (text stripped for privacy)
markets	Market Metadata	Market questions, categories, resolution info
ohlcv	Market OHLCV	Price/volume time series

Usage

from datasets import load_dataset

# Load a task split
ds = load_dataset("mlsys-io/EventXBench", "t1")
train_df = ds["train"].to_pandas()
test_df = ds["test"].to_pandas()

# Load market metadata
markets = load_dataset("mlsys-io/EventXBench", "markets")

# Load OHLCV time series
ohlcv = load_dataset("mlsys-io/EventXBench", "ohlcv")

Data Fields

T1: Market Volume Prediction

• event_group_id (str): Cluster ID for the event group
• condition_id (str): Polymarket market condition ID
• question (str): Market question text
• category (str): Market category
• tweet_count (int): Number of tweets in the event cluster
• unique_user_count (int): Distinct authors
• burst_duration_hours (float): Duration of tweet burst
• max_author_followers (int): Max follower count in cluster
• interest_label (str): Target label -- high (>80th pctl), moderate (40th--80th), low (<40th)
• ... (see full schema in the dataset viewer)

T2: Post-to-Market Linking

• tweet_id (int): Twitter post ID
• tweet_text (str): Tweet text content
• market_id (str): Polymarket condition ID
• market_question (str): Market question text
• embedding_score (float): Semantic similarity score

T3: Evidence Grading

• tweet_id (int): Twitter post ID
• condition_id (str): Polymarket condition ID
• tweet (str): Tweet text
• market (str): Market metadata
• question (str): Market question
• final_grade (int): Evidence grade 0-5
• llm_grade (int): LLM-assigned grade
• llm_confidence (float): LLM confidence score

T4: Market Movement Prediction

• tweet_id (int): Twitter post ID
• condition_id (str): Polymarket condition ID
• price_t0 (float): Price at tweet publication time
• delta_2h (float): Absolute price change at 2h horizon
• direction_label (str): up, down, or flat
• magnitude_bucket (str): small, medium, or large
• confound_flag (bool): Whether confounding events were detected

T5: Volume and Price Impact

• tweet_id (int): Twitter post ID
• condition_id (str): Polymarket condition ID
• price_impact_json (dict): Max absolute deviation from p0 at horizons (15m, 30m, 1h, 2h, 6h)
• volume_multiplier_json (dict): Total volume / 24h baseline at multiple horizons
• decay_class (str): transient, sustained, or reversal
• confound_flag (bool): Whether confounding events were detected
• Metrics: Spearman rho for price_impact and volume_multiplier, decay macro-F1

T6: Cross-Market Propagation

• tweet_id (int): Twitter post ID
• primary_condition_id (str): Primary market condition ID
• label (str): no_effect, primary_mover, propagated_signal
• sibling_count (int): Number of sibling markets
• moved_sibling_count (int): Number of siblings that moved
• primary_delta_h (float): Primary market price change
• confound_flag (bool): Whether confounding events were detected

Posts (Tweet Metadata)

• tweet_id (int): Twitter post ID
• text (null): Set to NULL for privacy -- use Twitter API for rehydration
• Additional metadata fields (timestamps, user IDs, etc.)

Privacy and Ethics

• Tweet text: Stripped from the public release per Twitter/X ToS. Tweet IDs are provided for authorized rehydration.
• Market data: Polymarket data is publicly available on-chain and included under fair use for research.
• No PII: User-level features are aggregated; no individual user profiles are released.

License

CC BY-NC 4.0