PXIN/fracture-crypto-dollar-bars

Fracture Crypto Dollar Bars

What This Is

This dataset contains Dollar Value Bars (also called USDT Bars) derived from raw tick-level trade data from Binance Futures USD-M perpetual contracts. The raw trades were transformed from time-based sampling to volume-based sampling, producing a more statistically robust representation of market activity.

Source Data

• Exchange: Binance Futures (USD-M Perpetuals)
• Pairs: BTCUSDT, ETHUSDT, SOLUSDT, BNBUSDT, HYPEUSDT, TRXUSDT
• Date Range: September 2025 — March 2026 (7 months)
• Raw Schema: id, price, qty, quote_qty, time, is_buyer_maker

Why Dollar Bars?

Traditional time-based bars (1-minute, 5-minute candles) sample the market at fixed clock intervals. This creates a fundamental statistical problem: low-activity periods are oversampled (producing noise) while high-activity periods are undersampled (missing information). The resulting return series exhibit conditional heteroskedasticity, serial correlation, and heavy tails — all of which violate the i.i.d. assumptions required by most machine learning algorithms.

The Research Foundation

The superiority of volume-based and dollar-value-based sampling over time-based sampling is well-established in the quantitative finance literature:

Marcos López de Prado (2018), Advances in Financial Machine Learning (Wiley):

"Sampling a bar every time a pre-defined amount of volume is exchanged leads to a process with superior statistical properties compared to time-based bars... Time bars oversample information during low-activity periods and undersample information during high-activity periods."

López de Prado identifies three primary information-driven bar types:

1. Tick Bars — sample every N trades
2. Volume Bars — sample every N units of volume
3. Dollar Bars — sample every N units of notional value

Dollar bars are preferred over tick and volume bars when analyzing instruments whose prices change significantly over time, as they naturally account for inflation and price drift.

Andersen, Bollerslev, Diebold & Labys (2003), Modeling and Forecasting Realized Volatility: Demonstrated that realized volatility estimators constructed from high-frequency returns suffer from microstructure noise. Sampling by business time (activity-based) rather than calendar time reduces this noise substantially.

Clark (1973), A Subordinated Stochastic Process Model with Finite Variance for Speculative Prices: Introduced the concept of "subordinated processes" — the idea that price evolution should be indexed by trading activity (volume/transactions) rather than clock time. This is the theoretical ancestor of all volume-bar methodologies.

O'Hara (1995), Market Microstructure Theory: Established that the trading process itself carries information. Clock-time sampling ignores the irregular spacing of transactions, discarding the informational content contained in the timing and intensity of trades.

Why This Matters for Machine Learning

Property	Time Bars	Dollar Bars
Sampling basis	Clock time	Trading activity
Returns distribution	Heavy-tailed, skewed	Closer to Gaussian
Serial correlation	High	Reduced
Heteroskedasticity	Pronounced	Attenuated
Inflation adjustment	None	Automatic
Regime representation	Uniform	Activity-proportional

For gradient-boosted tree models (LightGBM/XGBoost) and neural networks trained on tabular market data, dollar bars provide a more stable, better-conditioned feature space. The model learns from what the market actually did rather than what the clock happened to capture.

Bar Formation Logic

This dataset uses true local-reset dollar bars, not the common global-cumulative approximation:

1. Trades are streamed in chronological order (sorted by time)
2. quote_qty (the exact USDT value of each trade) is accumulated into a running bucket
3. When the bucket crosses the pair-specific threshold, a bar is emitted containing all trades that contributed
4. The bucket resets to zero and accumulation begins for the next bar
5. Trades are treated as atomic — they are never split across bar boundaries

Thresholds Per Pair

Pair	Threshold	Rationale
BTCUSDT	$1,000,000	Deepest liquidity, largest trades
ETHUSDT	$750,000	High liquidity
SOLUSDT	$250,000	Medium liquidity
BNBUSDT	$200,000	Medium liquidity
HYPEUSDT	$100,000	Lower float, thinner liquidity
TRXUSDT	$50,000	Low price per token

Schema

Column	Description
pair	Trading pair identifier
open	Price of first trade in bar
high	Highest price in bar
low	Lowest price in bar
close	Price of last trade in bar
volume_usdt	Total USDT value traded in bar (≥ threshold)
token_volume	Total token quantity traded in bar
cvd	Cumulative Volume Delta — aggressive buy volume minus aggressive sell volume
tick_count	Number of individual trades in the bar (volatility proxy)
open_time	Timestamp (ms) of first trade
close_time	Timestamp (ms) of last trade
return	Log-return from previous bar close (close / close_prev - 1)
price_impact	Return normalized by volume (return / volume_usdt)
cvd_delta	Change in CVD from previous bar
bar_duration_ms	Time span of bar in milliseconds
range_pct	(high - low) / open
body_pct	(close - open) / open
trade_intensity	Alias for tick_count

Cumulative Volume Delta (CVD)

CVD measures order flow imbalance:

• is_buyer_maker = false → buyer is taker (aggressive buy) → adds to CVD
• is_buyer_maker = true → buyer is maker (aggressive sell) → subtracts from CVD

Positive CVD = more aggressive buying. Negative CVD = more aggressive selling.

Use Cases

This dataset is suitable for:

• Market microstructure research — analyzing trade flow, liquidity dynamics, and price discovery
• Feature engineering for ML models — the bar structure naturally produces informative features (tick count as volatility proxy, CVD as flow metric, bar duration as activity measure)
• High-frequency trading strategy backtesting — dollar bars eliminate time-sampling bias
• Regime detection — bar frequency and duration encode market state information

What This Dataset Is Being Used For

This data powers a deterministic crypto perpetual trading system built on the following architecture:

1. Base Signal Engine — LightGBM gradient boosted trees trained on dollar bar features to generate directional alpha
2. Policy Engine — One-Step Flow Q-Learning (OFQL), an offline reinforcement learning policy that generates continuous actions [position, leverage] in a single deterministic forward pass
3. Execution — ONNX-compiled models running on a Rust hot path with zero-copy deserialization and TensorRT inference

The dollar bar format was chosen specifically because it provides the most robust tabular representation for gradient boosted trees and ensures training-serving parity between the Python research environment and the Rust production engine.

Files

• BNBUSDT_dollar_bars.parquet — 887,322 bars
• BTCUSDT_dollar_bars.parquet — 2,880,416 bars
• ETHUSDT_dollar_bars.parquet — 3,855,752 bars
• HYPEUSDT_dollar_bars.parquet — 904,006 bars
• SOLUSDT_dollar_bars.parquet — 2,965,440 bars
• TRXUSDT_dollar_bars.parquet — 324,408 bars

Citation

If you use this dataset, consider citing the foundational work on dollar bars:

López de Prado, M. (2018). Advances in Financial Machine Learning. Wiley.

License

Data derived from Binance public trade history. Provided as-is for research purposes.