Pltrr/btc15m-market-microstructure

BTC15M Market Microstructure

This dataset is a research collection for Polymarket's 15-minute Bitcoin UP/DOWN markets. It aligns public Polymarket market and order-book observations with point-in-time Bitcoin market context from Binance and reference-price data from Chainlink-related collection pipelines.

The package is designed to answer questions such as:

• How do YES and NO prices react when Bitcoin moves USD 10, 20, 50, 100, or more above or below the market reference price?
• How do spread, depth, order-book imbalance, order flow, and implied probability change as expiry approaches?
• At which time-to-expiry checkpoints are Bitcoin direction and Polymarket repricing most closely related?
• Could a five-share taker order have been filled from the displayed book, and what would its volume-weighted price have been?
• How do volatility, time of day, feed interruptions, reference-price ambiguity, and market-data quality affect a replay or model result?

It is a market-research dataset, not a profitable strategy, trading signal, or proof that an order would have filled live.

Package Summary

The current research package contains three active components:

Component	Parquet files	Rows	Bytes
Main sanitized market data	143	399,185,137	10,168,030,156
Historical extension	870	48,040,268	440,659,091
Research-ready add-on	6	265,065	10,105,904
Combined	1,019	447,490,470	10,618,795,151

The observed periods run approximately from June 24, 2026 through August 26, 2026. Coverage is not continuous. The high-resolution BTC15M observation layer is concentrated between July 31 and August 26, with documented internet and collector interruptions. Do not infer continuous coverage from the minimum and maximum timestamps alone.

What Is Included

Direction-grade data

Point-in-time inputs for direction, regime, calibration, and reaction studies:

• Binance BBO and public trades
• Bitcoin returns, realized volatility, spot/perpetual basis, funding, and open interest context where available
• Chainlink/reference prices and boundary-confirmation metadata
• signed Bitcoin gap from the market reference price
• 1-second market timelines
• TTE checkpoints at 600, 330, 300, 240, 180, 120, 60, and 30 seconds
• gap-threshold events and post-event reactions
• settlement labels and market-level quality flags

Execution-candidate data

Identity-checked order-book and public trade observations that can be used as inputs to an execution replay:

• compact order-book events
• validated asset_target_v2 market/token identity
• two-sided 1-second execution timelines
• 30-second depth snapshots
• displayed bid/ask capacity and five-share VWAP calculations
• public Polymarket trade events

execution_candidate means that an observation passed the published data- quality checks. It does not mean a live order was guaranteed to fill. Queue position, network latency, order acknowledgement, cancellation latency, partial fills, fees, and adverse selection still need to be modeled explicitly.

Diagnostic-only data

Derived or legacy views useful for diagnostics and historical comparison. These rows may overlap canonical observations or contain weaker identity/coverage quality. They must not be treated as execution-grade evidence without applying the supplied quality metadata.

Metadata and labels

• public market, condition, and token identifiers
• market lifecycle and settlement outcomes
• reference-price confirmation and ambiguity fields
• collection incidents and coverage statistics
• market-level direction and execution quality flags

Research-Ready Tables

The research_ready_addon contains compact starting points for analysis:

Table	Rows	Primary use
market_quality_flags	1,073	Filter markets by timeline, execution, reference, and rotation quality
collection_incidents_daily	83	Identify outages and feed-delivery anomalies
market_level_summary	1,100	Compare liquidity, spread, depth, volatility, and outcomes by market
tte_checkpoints	8,734	Point-in-time direction and repricing research by time to expiry
gap_reactions	188,358	Study YES/NO reaction after Bitcoin gap-threshold crossings
fillability_5share	65,717	Replay displayed five-share bid/ask capacity and VWAP

These tables are derived views and can overlap the lower-level event datasets. Do not add their row counts to raw event counts when estimating unique events.

Practical Uses

This dataset can support:

1. Direction and calibration research Build walk-forward UP/DOWN probability models using only point-in-time features. Evaluate Brier score, log loss, calibration error, and net expected value rather than accuracy alone.

2. Bitcoin-to-Polymarket lag analysis Measure how quickly YES/NO prices, spreads, and depth react to changes in Bitcoin gap, momentum, and volatility.

3. Time-to-expiry studies Compare market behavior at fixed TTE checkpoints and test whether a signal is useful early, mid-market, or near settlement.

4. Order-book microstructure analysis Analyze displayed liquidity, OBI, OFI proxies, microprice deviation, trade imbalance, complement-price deviations, and liquidity deterioration.

5. Execution replay research Estimate taker VWAP and displayed capacity for five-share orders, then add explicit fee, slippage, delay, queue, partial-fill, and adverse-selection assumptions.

6. Data-quality-aware backtesting Exclude interrupted, stale, identity-invalid, sequence-gapped, or reference-ambiguous periods instead of silently converting missing values to zero.

7. Complete-set and complement pricing research Study YES+NO bid/ask sums and whether an apparent pricing deviation survives executable depth, fees, fill risk, and legging risk.

Directory Layout

main_sanitized_market_data/
  direction_grade/
  execution_candidate/
  diagnostic_only/
  metadata_labels/

historical_extension/
  direction_grade/
  execution_candidate/
  diagnostic_only/
  metadata_labels/

research_ready_addon/
  direction_grade/
  execution_candidate/
  metadata_labels/

Use CURRENT_PACKAGE_INDEX.json as the source of truth for active components, file counts, row counts, byte totals, and superseded staging packages.

Example With DuckDB

The repository can be downloaded with huggingface_hub or queried after a selective snapshot download.

from huggingface_hub import snapshot_download
import duckdb

root = snapshot_download(
    repo_id="Pltrr/btc15m-market-microstructure",
    repo_type="dataset",
)

tte = f"{root}/research_ready_addon/direction_grade/tte_checkpoints/*.parquet"

result = duckdb.sql(f"""
    SELECT
        target_tte_sec,
        resolution_outcome,
        count(*) AS observations,
        avg(direction_signed_gap_usd) AS avg_signed_gap_usd,
        avg(yes_spread) AS avg_yes_spread,
        avg(no_spread) AS avg_no_spread
    FROM read_parquet('{tte}')
    WHERE direction_eligible
      AND quality_known
    GROUP BY 1, 2
    ORDER BY 1, 2
""").df()

print(result)

For large event families, query only the required date partitions and columns. Downloading the complete repository is not necessary for every analysis.

Data-Quality Rules

• Missing or unknown values remain NULL; they are not replaced with zero.
• Use direction_eligible for direction studies and execution_eligible for execution replays.
• Require mapping_version=asset_target_v2 for execution-sensitive work.
• Check timeline completeness, maximum gap, sequence gaps, stale-book flags, reconnect epochs, and Chainlink/reference confirmation before using a market.
• Settlement outcomes must only be joined after market resolution. They are labels, not point-in-time features.
• Historical legacy order-book data can be useful for direction/regime research but is not automatically execution-grade.
• Derived views overlap lower-level observations. Deduplicate with canonical market/time/token keys when combining families.

Known Limitations

• Coverage contains internet and collector interruptions and is not a perfectly continuous calendar sample.
• Displayed order-book depth is not proof of fill and does not reveal exact queue position.
• Binance and Chainlink/reference prices are related but not interchangeable; small boundary differences can change close-to-zero outcome interpretation.
• The package contains observational data, so model selection and repeated hypothesis testing can create false discoveries.
• Historical and current components have different execution-quality levels.
• No claim of profitability, causal predictability, or live-trading suitability is made.

Privacy and Safety

The publication package excludes:

• wallet or account data
• real orders, fills, positions, intents, and redeem records
• API keys, credentials, connection strings, and private keys
• raw SQL database dumps
• raw JSON/debug payloads and local filesystem paths

The publication manifests report zero sensitive findings for the active components. Users should still apply their own review before redistribution.

Publication Status

The package passed the documented privacy and credential audit. No license is asserted or implied. Repository visibility does not grant rights beyond those available under the applicable upstream Binance, Polymarket, and reference-data terms. Users are responsible for reviewing those terms before redistribution or commercial use.

This dataset is provided for research and engineering purposes. It is not financial advice. See SECURITY_AND_PRIVACY.md for the audit scope.