Information-Driven Bars

The core mechanism samples bars when the cumulative order flow imbalance reaches a predefined threshold. Imbalance is calculated as the difference between buyer-initiated and seller-initiated volume, normalized by total volume. When the absolute imbalance exceeds the threshold, a new bar is formed.

• Tick Rule: Classifies trades as buys if price ticks up, sells if price ticks down
• EMA of Imbalance: Uses an exponentially weighted moving average to smooth the imbalance signal
• Dynamic Thresholds: Thresholds can adapt to volatility regimes, creating more bars during high-activity periods

A variant where bars are sampled when the cumulative dollar value traded reaches a fixed threshold. This normalizes for the economic significance of transactions rather than raw share count, making bars comparable across assets with vastly different price levels.

• Notional Value: Each bar represents a constant amount of capital exchanged
• Volatility Normalization: Naturally adjusts for price level changes over time
• Cross-Asset Comparability: Enables consistent feature engineering across a multi-asset universe

These bars sample based on the imbalance of tick counts rather than volume. Each trade is assigned a signed tick (+1 for buy, -1 for sell), and a new bar forms when the cumulative signed tick count crosses a threshold.

• Equal Weight per Trade: Treats all trades equally regardless of size
• Microstructure Sensitivity: Captures information from order flow fragmentation
• Noise Reduction: Less sensitive to large block trades that may distort volume-based metrics

A specialized bar type that samples when the sequential run of same-direction trades reaches a threshold. Run bars capture the persistence of order flow direction, isolating periods where informed traders dominate one side of the market.

• Run Length: Counts consecutive trades in the same direction
• Informed Flow Detection: Long runs often indicate institutional order execution
• Threshold Calibration: Expected run length under a null hypothesis of random order flow determines the sampling threshold

Information-driven bars produce time series with superior statistical properties compared to time bars. Returns exhibit closer-to-normal distributions, reduced serial correlation, and more consistent variance.

• Homoscedasticity: Bar-level variance is more stable across the sample
• Reduced Autocorrelation: Mitigates the microstructure noise that plagues time-based sampling
• IID Approximation: Bars more closely approximate independent and identically distributed observations, a key assumption for many ML models

Practical deployment requires careful handling of tick data quality, threshold calibration, and computational efficiency. The sampling process must run in real-time for live trading or efficiently over historical archives for backtesting.

• Data Cleaning: Requires robust handling of trade corrections, cancellations, and exchange timestamps
• Threshold Selection: Typically calibrated to produce a target number of bars per day, often matching the average count of a comparable time frequency
• Streaming Architecture: Implemented as a stateful accumulator that emits a bar when the condition triggers

The core signal that drives information-driven bar sampling. OFI quantifies the net aggressive pressure in the market by measuring the difference between buyer-initiated and seller-initiated volume at the best bid and ask prices. A high positive OFI indicates strong buying pressure, while a negative OFI signals selling pressure. Information-driven bars trigger a new bar when the cumulative OFI exceeds a predefined threshold, ensuring bars are sampled only when meaningful directional information arrives.

The high-frequency random variation in prices caused by operational frictions like bid-ask bounce and order flow fragmentation. Time-based bars (e.g., 1-minute candles) sample this noise uniformly, diluting the signal. Information-driven bars are explicitly designed to filter out microstructure noise by only sampling when the informational content—measured by imbalance—is statistically significant. This produces a cleaner, more predictive time series for downstream models.

Information-driven bars come in three primary flavors, each using a different imbalance metric:

• Tick Imbalance Bars: Sample based on the imbalance between the number of buy and sell ticks.
• Volume Imbalance Bars: Sample based on the imbalance of buy and sell volume.
• Dollar Imbalance Bars: Sample based on the imbalance of buy and sell dollar value traded. Dollar bars are often preferred for long-horizon analysis as they account for corporate actions and inflation, while tick bars are most responsive to pure order flow dynamics.

A self-exciting point process that models the clustering of market events. The Hawkes process captures the empirical reality that trades and order book updates arrive in bursts, not uniformly. Information-driven bars can be seen as a deterministic, threshold-based alternative to the stochastic Hawkes framework. Both approaches share the same goal: identifying periods of heightened market activity driven by endogenous information cascades rather than exogenous noise.

The statistical properties of financial time series evolve over time, causing predictive models to degrade. Information-driven bars provide a natural defense against concept drift because they adapt their sampling frequency to the current market regime. In high-volatility, high-information periods, bars form rapidly. In low-volatility, quiet markets, bars form slowly. This self-adjusting property ensures that models trained on these bars see a more stationary distribution of features.