Bar Build

The State Engine produces 9 bar types in two families: activity-sampled (triggered by volume, trade count, or time) and information-driven (triggered by statistical conditions). Each bar type has init_* and update_* functions in the Numba kernel following the pattern: receive current state + new trade → (next_state, is_emitted, bar_tuple).

Activity-Sampled Bars

Sample the market at predetermined thresholds. Produce more observations when activity is high, fewer when it's quiet.

Bar Type	Trigger	Default Param	Kernel File
Tick	`trade_count >= count`	`count: 500`	`kernel/bars/tick_bar.py`
Volume	`cumulative_volume >= threshold`	`threshold: 10000.0`	`kernel/bars/volume_bar.py`
Dollar	`cumulative_dollar_vol >= threshold`	`threshold: 500000.0`	`kernel/bars/dollar_bar.py`
Time	`elapsed >= interval_seconds`	`interval: 30.0s`	`kernel/bars/time_bar.py`

Tick, volume, and dollar bars adapt to market tempo. Dollar bars additionally account for price changes — a bar at $50k BTC takes fewer trades than at $20k BTC.

Information-Driven Bars

Trigger on statistical conditions rather than fixed thresholds. Each maintains an exponentially weighted running estimate of the expected threshold. Inspired by modern financial ML for producing IID-like samples from non-stationary data.

Bar Type	Trigger	Default Param	event_value
Imbalance	abs(cum signed volume) >= K	`k: 100.0`	Signed volume at trigger
Run	Consecutive same-direction trades >= W	`window: 10`	Run length
Volatility	EWMA variance > threshold (n > 2)	`threshold: 0.0001`	EWMA variance
Dollar Imbalance	abs(cum signed dollar vol) >= K	`k: 500000.0`	Signed dollar volume
CUSUM	s_pos >= h or s_neg <= -h	`threshold: 0.5`	CUSUM value

Trigger Details

Imbalance Bars: Track cumulative signed volume (buy=+vol, sell=-vol). Trigger when absolute imbalance exceeds K. Captures order flow toxicity — large directional pressure signals informed trading.
Run Bars: Count consecutive same-direction trades. Trigger when run length exceeds W. The kernel also supports directional variants (buy-run, sell-run) that count only one side.
Volatility Bars: EWMA of squared returns. Trigger when running variance exceeds threshold. Fire during volatile periods, stay quiet during calm.
Dollar Imbalance Bars: Dollar-volume-weighted imbalance — large-dollar trades carry more weight.
CUSUM Bars: Dual-sided filter — s_pos and s_neg track deviation from a drift term. When either crosses the threshold, a bar is emitted and both reset. Detects structural breaks.

Trade Signing

Trade direction is inferred via a priority chain of signing methods (configurable via sign_methods):

DSIDE — use exchange-reported aggressor flag if available
QUOTE_INFER — compare trade price to mid-price; buy if trade ≥ ask, sell if trade ≤ bid
LEE_READY — quote test + tick test fallback

The p_buy column stores the continuous probability (0.0–1.0) that a trade was buyer-initiated, enabling signed_volume = size * (2*p_buy - 1) as a continuous alternative to discrete +1/-1 direction.

Bar Output Schema

Every bar has 13 columns in the kernel output (plus metadata columns added by the writer):

Column	Description
`start_time`	Bar opening timestamp
`end_time`	Bar closing timestamp
`open`	Opening price
`high`	Highest price in bar
`low`	Lowest price in bar
`close`	Closing price
`volume`	Total traded volume
`dollar_volume`	Total traded notional
`trade_count`	Number of trades
`vwap`	Volume-weighted average price
`avg_trade_size`	Average size per trade
`bar_index`	Sequential bar number
`event_value`	Trigger value (0 for activity-sampled, trigger-specific for information-driven)

param_set_id strings encode configuration parameters (e.g. "k_100", "t_0.0001", "w_10") to distinguish bars generated with different thresholds.

Configuration

state_engine:
  micro_batch_size: 200000
  bar_groups:
    - name: liquid_equities
      symbols: [SPY, QQQ, NVDA, AAPL]
      bars:
        - type: time
          interval_minutes: 1
        - type: tick
          count: 200
        - type: dollar
          threshold: 100000.0
        - type: volume
          threshold: 10000.0
        - type: imbalance
          k: 100.0
        - type: run
          window: 10
        - type: volatility
          threshold: 0.0001
        - type: dollar_imbalance
          k: 100000.0
        - type: cusum
          threshold: 0.5
      snapshots:
        period_seconds: 60.0
        depth_levels: 10
      trade_signing:
        method: quote

Bar groups allow per-symbol bar configurations — different instruments get different thresholds based on their liquidity profiles.

The active_bar_types config field (default: ["tick", "volume", "dollar", "time"]) controls which bar types the kernel actually computes. Only listed types allocate accumulator arrays and emit output. Set it explicitly to include information-driven bars in production.

Batchers

Two batcher implementations feed data to the State Engine:

Batcher	Source	Use Case
MockBatcher	Synthetic NumPy arrays (60% trades, 40% LOB)	Testing and benchmarking — no database needed
SourceBatcher	`StateEngineReader` over a real (venue, symbol) pair	Production — streams RecordBatches from the target engine

Activity-Sampled Bars​

Information-Driven Bars​

Trigger Details​

Trade Signing​

Bar Output Schema​

Configuration​

Batchers​