Free lesson · HFT & microstructure

HFT research promotion, drift and operating limits

Open interactive lessonPractice calculationsExplore labs

Split by sessions or longer contiguous periods, preserving distinct volatility and liquidity conditions. Hold out days for final evaluation; do not mix event-level samples from the same session across training and test. Many closely spaced events share the same latent market state.

Monitor latency quantiles, reject rates, fill ratios, signed markouts, realised versus forecast volatility and inventory concentration. Distinguish data drift from concept drift: a feature distribution may change without losing predictive value, while the same feature values may acquire a different outcome relationship.

Use shadow decisions and constrained paper operation to compare intended orders with plausible fills. Keep independent stale-feed, position, loss and order-rate limits outside the learned policy. A malfunctioning learner should not decide whether its own risk controls apply.

Apply it in a strategy

• Freeze a promotion dossier with candidate policy, costs, queue assumptions, training data and final holdout.
• Monitor per-session diagnostics against uncertainty bands estimated from comparable historical blocks.
• Use predetermined rollback and retraining criteria; any new parameter choice starts a new documented research cycle.

Research deliverable

Prepare a shadow-trading report that reconciles forecasts, desired orders, plausible fills, costs and operating failures by session.

Research checkpoint · reviewed 11 September 2026

These sources inform the questions to test. A result is conditional on its data, simulator and evaluation design. The examples in this module are teaching calculations, not reproductions of the reported experiments.

Regime and inventory failure modes. This newly submitted preprint studies a distributional-DQN market maker. The setup, stress discussion and limitations were reviewed in full text. Evaluation uses a zero-intelligence simulated order book, unit lots and at most one order per side; there is no historical trading sample. Persistent directional flow can saturate inventory, and the authors study regime-aware fine-tuning. Our implementation lesson is to test inventory saturation and simulator assumptions, rather than treat simulated improvement as an executable market edge.

Further reading: Moret & Lillo · Deep Learning of Robust Market Making under Regime-Switching Order Flow · 10 September 2026 ↗

Review a causal volume-clock feature and its measurement limits →

Research sources, review dates and limitations

Extend the research question

Compare inventory management in a continuous-price instrument with binary terminal settlement. Test a quote policy under both adverse order flow and resolution risk.

Continue with the connected research module →

from math import sqrt

def fill_diagnostic(fills,attempts,reference):
    if attempts<=0 or not 0<=fills<=attempts or not 0<=reference<=1: raise ValueError("Invalid counts")
    p=fills/attempts
    se=sqrt(p*(1-p)/attempts)
    return p,se,(p-reference)/se if se>0 else None

print(fill_diagnostic(400,1000,.45))