Inside the ZunBit AI Engine: The Models, Signals and Guardrails Explained in Plain English
You deserve to know what’s inside the box. Here is a plain-English, no-fluff breakdown of the ZunBit AI engine — the models, the data, the guardrails and the kill-switch.
Step 1 — Raw data
We ingest tick-level price data for 1,400+ instruments across major exchanges from SOC-2 certified providers (Refinitiv, Polygon, ICE). We also pull macro feeds (central-bank calendars, CPI releases), news feeds (Bloomberg, Benzinga) and corporate-action feeds. Nothing scraped. Nothing bootlegged. All licensed.
Step 2 — Feature engineering
From the raw data we derive ~240 engineered features per instrument: volatility clusters, order-book pressure, cross-asset correlations, regime indicators, news sentiment, options skew and more. Each feature has a documented lineage — our quants can reproduce it from scratch.
Step 3 — The model ensemble
We run three models in parallel: a gradient-boosted base learner (fast, interpretable), a transformer-based sequence model (captures temporal patterns), and a Bayesian meta-learner that reconciles the two and outputs a signal, confidence score and attribution vector. Nothing proprietary we can’t explain — just well-engineered, well-tested machine learning.
Step 4 — Guardrails
Every candidate signal passes three gates before publication:
- Risk-cost gate: if predicted drawdown exceeds your cap, suppressed.
- Liquidity gate: if the position would move more than 10% of order-book depth, suppressed.
- Regime gate: if the current market regime isn’t represented in the training window, suppressed.
Step 5 — Live monitoring
In production, the ensemble is monitored against its backtest baseline. If live-vs-expected drift exceeds a pre-set threshold, the model is auto-retired and the ensemble reverts. Our research desk is paged. You see a note on your dashboard.
Step 6 — The kill switch
Every engine instance has a global kill switch that halts all new order placement across all accounts. It is triggered automatically by (a) data-feed failure, (b) broker-API degradation, (c) extreme volatility beyond the model’s trained regime, or (d) manual intervention by the on-call engineer. Published publicly at status.zunbitai.com.
What we don’t do
- No social-signal following. Twitter sentiment is noise 95% of the time.
- No reinforcement-learning agents in production. Too unpredictable in tails.
- No price-action-only strategies. We require a multi-feature prior.
- No crypto. Our risk-engine design assumes regulated venues.
If we can’t explain it, we don’t ship it. If we can’t audit it, we don’t trust it.
Published October 2026 by Matteo Ferraro, ZunBit AI CTO. Not investment advice.