Our own model. It learns from its mistakes — carefully.
AlgoNexus runs its own open-weights language model, fine-tuned for the markets we serve. Quantitative models, a portfolio optimiser and a deterministic risk engine sit around it, and the model can never place an order on its own.
Automated learning. Gated releases.
Daily returns are noisy, so we automate data, training and evaluation — but a new model goes live only after it proves itself out of sample. Read the AI transparency statement for its limits.
- 01New dataFilings, results, regulator circulars and market data.
- 02Quality and licensing checksBad or unlicensed data is dropped before training.
- 03Outcome labelsWhat actually happened after our own past predictions.
- 04Fine-tuneLoRA / preference tuning on on-demand GPUs.
- 05Evaluation gatesMust beat the current model and simple baselines, net of costs, with no worse drawdown.
- 06Shadow runRuns on paper alongside production before any user sees it.
- 07Human approvalA named reviewer signs off the release.
- 08Canary with rollbackA small share first; automatic rollback on regressions.
Layers of the engine
Research model
Reads filings, results, circulars and news; produces structured, cited analysis.
Quant models
Momentum, volatility, factor exposure and regime estimates with calibrated probabilities.
Portfolio optimiser
Turns signals and your policy into position sizes and rebalances.
Decision policy
Combines signals, confidence and disagreement; abstains when evidence is weak.
Risk gate
Deterministic, model-independent, fail-closed checks on every order.
Execution quality
Learns slippage and fill behaviour per broker and venue.
- ✓Kill switchoff
- ✓Market hoursNYSE open
- ✓Quote freshness0.6 s ≤ 5 s
- ✓Buying power$6,420 ≥ $5,174
- ✓Position cap8.9% ≤ 10%
- ✓Cash reserve6.1% ≥ 5%
- ✓Duplicate ordernone in 60 s
- ✓Policy versionv7 (signed)
Start on paper. Go live only when you’re ready.
Set a goal, see whether it’s realistic, and watch AlgoNexus work with simulated money first. No card needed for the 14-day trial.