Architecture Specification
BitgoAI's stack decomposes into four operational layers — data, model, execution, and risk control — connected through a continuous feedback loop where strategy performance retraining the upstream models. The structural pattern is conventional production ML. The differentiator is the claimed sub-millisecond execution latency, which implies hardware acceleration colocated with venue matching engines.
Input vectors span market price behavior, order flow microstructure, capital movement paths, and sentiment data. The first three are standard. Sentiment introduces noise risk; the backtest burden falls on proving signal-to-noise ratio exceeds execution costs. No feature importance, signal decay curves, or training window specifications are disclosed in the release.
Risk Control as a Deterministic Variable
The risk layer operates continuously across volatility regimes, strategy behavior, and capital flow. The announcement claims automatic adjustment capability: position control, strategy degradation, and trading restrictions under abnormal conditions. This is the most consequential claim and the least substantiated.
Latency-sensitive systems without robust circuit breakers fail in cascades. No drawdown thresholds, kill-switch latency, or stress-test results are disclosed. The control logic described is plausible but unverified; a risk-adjusted verdict cannot be issued on architectural diagrams alone.
What to Track
Three metrics warrant measurement before any capital allocation decision:
- Audited Sharpe ratio across a minimum 24-month window including a drawdown period.
- Realized execution latency measured at the venue co-location point, not internal benchmarks.
- Slippage distribution under stressed liquidity conditions, segmented by volatility regime.
Sub-millisecond execution claims are common in vendor literature. The determining factor is measurement methodology. Until those figures are disclosed, the infrastructure thesis remains an unverified specification, not a track record.