Adapts in production. Never falls to the same regime twice.
Single-strategy trading bots that die the first time the regime changes against them. Static allocations that need a human to rebalance after the loss has already happened. Black-box ML agents that learn but can't show their work. Mahoraga's defining property is the opposite: every loss, every miscall, every drawdown is observed and shifts the next decision, and every shift is on the ledger.
Mahoraga is named for the adaptive entity that cannot be killed by the same technique twice. The product carries the same property. A Gaussian-mixture regime classifier reads the market state in real time; a LinUCB contextual bandit selects the strategy arm best-suited to that regime given current context; an append-only adaptation ledger records every regime call, every arm switch, every parameter update. The regime that hurt the agent yesterday tilts its posterior away from the strategies that hurt in that regime. The adaptation is online, the exploration is confidence-bounded, and every decision is on the ledger.
The current market sits in one of four GMM-identified regimes (highlighted in gold below). The bandit's posterior for each regime / arm pair is updated by every closed trade, losses in a regime push the bandit away from the arms that produced them, the next time that regime is detected.
Every closed loss updates the bandit's posterior for the regime / arm pair that produced it. The same arm that lost in a given regime is sampled less aggressively the next time that regime is detected. The agent does not get hit by the same combination twice.
A Gaussian-mixture model identifies the prevailing market regime from a rolling 60-day feature window, realised vol, cross-sectional dispersion, term structure, correlation collapse. Re-fit daily; re-classified intraday.
For the identified regime, a contextual bandit selects the best-performing arm given current context. Exploration is bounded by an upper-confidence interval that widens during regime transitions and tightens as the agent accumulates evidence.
Every regime call, every arm switch, every weight update is hash-chained and exportable. Tamper-detectable on export; seven-year retention by default; PDF and CSV out.
Per-arm drawdown caps, per-regime exposure ceilings, post-loss cooldown windows. The agent governs its own behaviour, there is no separate operator to override the controls mid-run.
Initial library covers FX trend, gold mean-reversion, equity-index momentum, and a cash-defensive sleeve. Each arm implements a small Python protocol and runs through the same risk pipeline, no special privileges, no special exits.
Mahoraga is configured in Python. Strategy arms are pluggable; the risk controls and adaptation ledger are not.
# mahoraga.config.py
from mahoraga import (
GMMRegimeClassifier,
LinUCBAgent,
RiskControls,
)
classifier = GMMRegimeClassifier(
n_regimes=4,
feature_window="60d",
refit_cadence="1d",
features=["realised_vol", "dispersion",
"term_structure", "corr_collapse"],
)
agent = LinUCBAgent(
classifier=classifier,
strategy_arms=[
"fx_trend",
"gold_mr",
"idx_mom",
"cash_defensive",
],
confidence_alpha=1.0, # exploration bound
risk=RiskControls(
per_arm_drawdown_cap=2.0, # % of NAV
post_loss_cooldown="30m",
max_exposure_per_regime=15.0, # % of NAV
),
)
agent.run()Tell us about your venues, your strategies, and the governance gap you're trying to close. We respond within two business days, or not at all.