Quantitative machine learning · 2025—26
GAIA Organism
An autonomous market-research organism that only acts when the evidence clears the gate.
Active milestone M3
The project
A triple-barrier directional outcome specialist for BTC and Gold, combining event-driven sampling, conformal prediction, and strict temporal validation.
The hard part
Financial time series carry overwhelming tick noise, regime shifts, and overlap leakage. Conventional predictive models can look accurate in backtests and decay quickly once exposed to live market structure.
The response
GAIA treats a forecast as a conditional evidence problem. CUSUM-filtered events reduce the stream to meaningful structural changes; purged and embargoed walk-forward validation protects the temporal boundary; conformal bounds decide whether a signal is allowed through.
System architecture
Built to make the decision legible.
- 01
CUSUM event detection filters stationary tick noise before any model sees a sample.
- 02
Purged and embargoed validation removes overlap leakage from the training and evaluation windows.
- 03
Quarter-Kelly allocation converts passing evidence into a bounded position rather than an unbounded bet.
- 04
Memory-pinned ROCm inference on AMD MI300X accelerators targets sub-15ms response.
What the system is meant to deliver
Outcomes and operating constraints.
- Signals publish at least 30 minutes ahead of structural volatility shifts.
- The model targets greater than 90% directional precision rather than optimizing accuracy alone.
- Purged cross-validation prevents temporal lookahead contamination.
- Memory-pinned ROCm inference targets response below 15ms on AMD MI300X.
PyTorch · ROCm · AMD MI300X · CUSUM filtering · Conformal prediction