FreqAI is the prediction box. The layers around it are what we actually built.
CORE
FreqAI × LightGBM, ~130 pairs
5-minute candles on MEXC spot, models retrained every 12 hours per pair.
The target is a triple-barrier classification — which gets hit first, the +1% take-profit
or the -1% stop — and the model is regularized after a 677,729-point calibration audit
caught it memorizing instead of predicting.
CORE
Dual-LLM advisory gate
A local gemma model tags the market regime every hour from price momentum plus
volume, volatility and cross-pair dispersion; Claude writes a risk directive three
times a day, re-evaluated instantly (both directions) when the regime flips. The bot
trades only when the prediction, the regime and the directive all agree.
NEW
AI post-mortems + news cross-check
Every closed trade is autopsied by an LLM against measured facts (max adverse
excursion, pre-entry pump, exit reason) — now with the pair's recent news attached,
so it can tell an unlucky accident from a foreseeable loss. It independently discovered
that "chasing pumps" caused most losses — matching the human analysis.
NEW
Nightly LLM researcher
Every night Claude reads the bot's own performance dossier and proposes testable
hypotheses in a constrained DSL. Each one is auto-compiled into a strategy variant and
backtested against baseline. Only proven winners are surfaced for human approval.
NEW
Per-pair news: collect, store, quarantine
Hourly news sweeps, plus a targeted deep-dive every 6 hours on held and
recently-losing pairs. Facts are stored in SQLite with expiry timestamps; hack or
delisting headlines auto-quarantine the pair for 24h and show as red cards on the
dashboard. It caught BONK's $20M governance-attack drain on day one.
NEW
Execution honesty
A rate limiter keeps unblock moments from bursting into correlated positions
(max 3 new entries per 20 minutes), and stake size is capped at 0.1% of each pair's
24h volume. When backtests revealed most paper profit came from full-size fills in
thin books, we chose honest numbers over flattering ones.
DISCIPLINE
Backtest kills bad ideas
On its first night the researcher proposed a plausible threshold change that
backtesting showed would have lost 7,000+ USDT (paper). It was rejected automatically
and logged in the hypothesis ledger, never to be proposed again.
