A repo worth a look: ahmetsbilgin/regime-aware-optimal-quoting, a Birkbeck MSc dissertation on extending Avellaneda-Stoikov to regime switching.

What it does

Four pieces:

1. Rolling-window LOB features: log-return, realized vol, relative spread, depth imbalance.
2. Gaussian HMM with sticky self-transition priors (EM, log-space forward-backward).
3. Regime-coupled HJB solved backward by ODE integration, one value function per regime with transitions via a generator matrix Q.
4. Online Bayesian filter for the current regime posterior, used to blend quotes at runtime.

What I found interesting

The HMM fits the problem: regime detection here is filtering, not classification. Sticky priors keep the posterior from flickering, which would otherwise destabilize the quotes downstream.

The HJB is regime-coupled rather than “A-S run per regime”: the backward ODE accounts for the chance of regime transitions within the horizon. If you are in a low-vol regime but likely to transition to high-vol, your current quotes already reflect it.

Caveats

The backtest simulator (src/backtest/simulator.py, src/execution/order_manager.py) uses naive crossing fills at top-of-book, unit size, no queue position, no partial fills. That is the main gap for a market-making backtest — adverse selection is modeled in the control problem but not in the fill simulation.

The README advertises walk-forward validation and grid search, but I did not find the harness in examples/ — there is a single-run run_backtest.py, no walk-forward loop. Possibly in a branch or not yet committed.

Things I would add

• Heavy-tailed emissions for the HMM.
• GPU port of the backward ODE: embarrassingly parallel across the inventory grid.
• Online HMM updates rather than a fixed historical fit.

Code @ github.com/ahmetsbilgin/regime-aware-optimal-quoting.