How ZunBit AI actually thinks — in words, not marketing.
If you’ve ever been told an AI “learns from millions of data points” without a shred of detail about what, how or why, this page is the antidote. Below is a direct, non-technical walkthrough of the ZunBit AI pipeline from raw tick to published signal.
1. Data ingestion
We ingest level-1 price data for 1,400+ instruments across equities, ETFs, commodities and major FX pairs from independent, SOC-2 certified providers (Refinitiv, Polygon.io and ICE Data Services). Every tick is timestamped, deduplicated and parked in cold storage before the model touches it. We also integrate macro and news feeds (Bloomberg Terminal, Benzinga Pro, FactSet) for context — not as a trading trigger but as a sanity check.
2. Feature engineering
We derive roughly 240 engineered features per instrument — volatility clusters, order-book pressure, cross-asset correlations, regime indicators, flow signals and anomaly scores. Every feature has a documented lineage so a quant on our team can reproduce it from scratch. No feature survives in production without passing a leakage test and an out-of-sample robustness test.
3. The model ensemble
ZunBit does not run a single model. It runs an ensemble: a gradient-boosted base learner, a sequence model for temporal patterns and a Bayesian meta-learner that reconciles their outputs. The meta-learner outputs a signal, a confidence score and a feature-attribution vector — which is what powers the “plain-English rationale” you see on each signal.
4. Guardrails
Before a signal ever reaches the dashboard it passes three guardrails:
- Risk cost gate. If predicted drawdown exceeds the user’s cap, the signal is suppressed.
- Liquidity gate. If order-book depth would require >10% market impact, suppressed.
- Regime gate. If the current market regime wasn’t represented in the model’s training window, suppressed.
5. Backtesting & walk-forward validation
Every model release passes walk-forward validation across at least 10 years of out-of-sample data, including the 2008, 2020 and 2022 drawdowns. We publish backtest reports to members monthly and keep the raw results in an audit log you can request.
6. Monitoring in production
Live model performance is monitored against its backtest baseline. If live-vs-expected drift exceeds a pre-set threshold, the model is automatically retired and the ensemble reverts to the previous version. Our research desk is paged. Nothing silent, nothing hidden.
7. What the AI will not do
It will not promise profit. It will not predict crashes. It will not generate signals for an instrument it hasn’t been validated on. It will not override your drawdown cap. And it will not fabricate reasoning — if the model can’t explain itself, we suppress the signal.
Trading intelligence is only worth what you can audit. ZunBit AI is built to be audited.