Research Log

2026-02-12: Trade-Only Baseline Exploration

Objective

• Improve event-driven strategy quality under:
• limit-only execution
• order latency = 0.3s
• historical trade-only backtest (no full L2/orderbook stream)

Added strategies

• baseline_regime_switch_micro
• baseline_flow_imbalance_pulse

Experiment A: 1-day slice

Command:

uv run atc report leaderboard \
  --symbol ETH_JPY \
  --start 2026-01-01 \
  --end 2026-01-01 \
  --strategy baseline_momentum_micro \
  --strategy baseline_mean_reversion_micro \
  --strategy baseline_regime_switch_micro \
  --strategy baseline_flow_imbalance_pulse

Result summary (sorted by alpha_sharpe):

• baseline_regime_switch_micro: sharpe -488.49, equity -171.97, turnover 2.02
• baseline_momentum_micro: sharpe -642.49, equity -263.33, turnover 3.42
• baseline_mean_reversion_micro: sharpe -645.02, equity -177.95, turnover 2.90
• baseline_flow_imbalance_pulse: sharpe -652.92, equity -242.67, turnover 3.18

Observation:

• baseline_regime_switch_micro reduced turnover and improved alpha sharpe versus existing momentum/reversion baselines on this day.

Experiment B: 7-day slice

Command:

uv run atc report leaderboard \
  --symbol ETH_JPY \
  --start 2026-01-01 \
  --end 2026-01-07 \
  --strategy baseline_momentum_micro \
  --strategy baseline_mean_reversion_micro \
  --strategy baseline_regime_switch_micro \
  --strategy baseline_flow_imbalance_pulse

Result summary (sorted by alpha_sharpe):

• baseline_regime_switch_micro: sharpe -646.75, equity -10130.27, turnover 105.08
• baseline_mean_reversion_micro: sharpe -707.21, equity -9459.71, turnover 117.24
• baseline_flow_imbalance_pulse: sharpe -708.73, equity -10801.95, turnover 120.87
• baseline_momentum_micro: sharpe -768.32, equity -15107.35, turnover 157.42

Observation:

• All tested trade-only baselines remain negative after latency-aware limit execution.
• baseline_regime_switch_micro is currently best by alpha sharpe and shows lower turnover.

Bottlenecks identified

• Trade-only historical data lacks orderbook microstructure needed by MM/queue imbalance class strategies.
• Order replacement is still frequent under high event rates.
• Signal-to-noise ratio is weak without stronger regime/time filters.

Next research steps

1. Add strategy-level holding/cooldown policies that target lower turnover under event bursts.
2. Introduce time-of-day gating (session-specific thresholds).
3. Build walk-forward splits and optimize only on train windows.
4. Start collecting forward WebSocket orderbook logs to unlock L2-aware models.

---

2026-02-12: Order Cadence Control (5-minute default)

Motivation

• Event-driven loop was submitting too frequently for practical operation.
• Desired operating style: continuous monitoring, but orders only every ~5 minutes to several hours.

Implementation

• Added --min-order-interval-sec to backtest and leaderboard commands.
• Default set to 300 seconds (5 minutes).
• Strategy evaluation still runs for every event; only new submissions are throttled.

Results (same windows as above, 4-strategy subset)

1-day (2026-01-01):

• orders dropped from 200-300 range to ~60-130
• best alpha sharpe improved from about -488 to -299
• best final equity improved from about -172 to -16

7-day (2026-01-01 to 2026-01-07):

• orders dropped from ~11k-17k to ~1.3k-1.6k
• alpha sharpe improved from about -647 to around -239
• final equity improved materially (still negative)

Interpretation:

• Main issue was not only signal quality; execution cadence was too aggressive.
• Throttling makes behavior closer to intended short-term style and reduces over-trading loss.