Ml4t Build Bars

Build Bars

Time bars sample by the clock, producing bars with wildly different information content - a 5-minute bar during the open contains 100x more trades than one at 2pm.

The Problem

Standard time bars (1-min, 5-min, daily) sample at fixed intervals regardless of market activity. During high-volume periods, a single bar compresses thousands of trades; during quiet periods, a bar may contain just a handful. This creates heteroscedastic returns that violate the i.i.d. assumptions of most ML models. Volume and dollar bars sample by activity instead, producing bars with roughly equal information content and returns closer to normality.

The Pattern

WRONG

import polars as pl

# Only using time bars - uneven information per bar
trades = pl.read_parquet("trades.parquet")
bars_5min = (
    trades.group_by_dynamic("timestamp", every="5m")
    .agg(
        open=pl.col("price").first(),
        close=pl.col("price").last(),
        volume=pl.col("size").sum(),
    )
)
# Bar at 9:30 has 5000 trades, bar at 14:00 has 50 trades

CORRECT

import numpy as np
import polars as pl

def bar_index(dollar_vol: np.ndarray, threshold: float) -> np.ndarray:
    """Bar id per trade: accumulate, close on the crossing, then RESET.

    Bucketing by the running total (cum_dollar // threshold) never resets, so
    an overshooting bar steals from the next: four $800 trades against a
    $1,000 threshold give $1,600, $800, $800, not two $1,600 bars."""
    idx = np.empty(len(dollar_vol), dtype=np.int64)
    bar, run = 0, 0.0
    for i, value in enumerate(dollar_vol):
        idx[i] = bar
        run += value
        if run >= threshold:
            bar, run = bar + 1, 0.0
    return idx

def build_dollar_bars(trades: pl.DataFrame, threshold: float) -> pl.DataFrame:
    """Build dollar bars: each bar contains ~threshold dollars traded."""
    trades = trades.sort("timestamp").with_columns(dollar_vol=pl.col("price") * pl.col("size"))
    return (
        trades
        .with_columns(bar_idx=bar_index(trades["dollar_vol"].to_numpy(), threshold))
        .group_by("bar_idx")
        .agg(
            timestamp=pl.col("timestamp").first(),
            open=pl.col("price").first(),
            high=pl.col("price").max(),
            low=pl.col("price").min(),
            close=pl.col("price").last(),
            volume=pl.col("size").sum(),
            dollar_volume=pl.col("dollar_vol").sum(),
            n_trades=pl.len(),
        )
        .sort("bar_idx")
    )

# Threshold: target ~same number of bars as time bars over the day
bars = build_dollar_bars(trades, threshold=1_000_000)  # $1M per bar

Choosing the Threshold

Calibrate so dollar bars produce roughly the same number of bars as time bars over the same period.

# Estimate: total dollar volume / desired number of bars
total_dollar_vol = (trades["price"] * trades["size"]).sum()
n_time_bars = 78  # 6.5 hours * 12 five-minute bars
threshold = total_dollar_vol / n_time_bars

Bar Type Comparison

Guardrails

• Dollar bars require tick-level trade data (timestamp, price, size) - cannot build from OHLCV
• Thresholds are symbol-specific: $1M/bar for AAPL vs $50K/bar for a small-cap
• Volume and dollar bars are not directly comparable across symbols - normalize returns
• Overnight gaps should be handled (exclude or flag the first bar of each session)

Production Implementation

No ml4t-* library covers bar construction. Use ml4t-data to fetch tick data, then apply the dollar-bar logic above. The canonical schema (timestamp, open, high, low, close, volume) from ml4t-data applies to the output bars.

Checklist

• Tick data available with timestamp, price, and size columns
• Threshold calibrated to produce reasonable bar count (~same as time bars)
• Bars include OHLCV, dollar volume, and trade count
• Overnight/session boundaries handled
• Returns closer to normal verified (Jarque-Bera test or QQ plot)