TradingAI 12.0 ~Nexus — Devlog

What happened since the last devlog: three models became a whole system and it’s running live.

What changed

The 8.0 system could predict, but it couldn’t trade. So it grew into a two-stage system:

Stage 1: 39 candle models. Instead of one timeframe, now 13 (from 1 minute to 1 day), each with Close, High/Low, and Volume variants. Every model looks at the last 200 candles and rolls its predictions up to 100 steps forward, the familiar ChatGPT principle, just on 13 time levels at once. Trained on 5 years of Binance data (~2.6M candles).

Stage 2: 6 signal models. The predictions from all timeframes feed into six classifiers (1m to 1d). Each learns a single question:

Will my take-profit or my stop-loss be hit first?

Answer: BUY / SELL / HOLD plus a confidence value that tells you how sure the AI is of itself.

**The server is running.**live Binance data, AI prediction in real time, TradingView-style chart, paper trading with fees and stats. The legacy of 8.0, just with 45 models behind it instead of three.

Still everything local, offline, no cloud, as it has always been.

Why so many SL/TP tests and timeframe signals?

Because the AI predicts direction it doesn’t tell us how far the move will go. And that’s where the real money is decided.

Why 100 SL/TP combinations. A tight target (+0.3%) is hit often but earns little; a wide one (+7%) earns a lot but rarely hits. The same signal can be profitable with one setting and ruinous with another. Winning often doesn’t mean making money ,the win rate only matters together with the risk/reward. So instead of guessing “maybe 1% is fine,” we simulate every combination on real price paths, subtract the fees, and let the numbers decide. No gut feeling, just math.

Why a separate signal per timeframe. Each timeframe is a different game. The 1m signal trades hundreds of times a day — lots of trades, small moves, high fee impact. The 1d signal trades rarely ,few trades, big moves, more risk per single decision. A setup that works on 5m can be a disaster on 4h. Every timeframe gets its own model and its own SL/TP test, so we find out independently which horizons actually have an edge.

What’s next

The first complete test run was too good to be true: +1.79% per trade at a 99.6% win rate. The analysis showed why: the models were trained on the entire history including the test period. The AI knew the answers. A classic leak.

The plan:

Retrain all models with a hard training cutoff, so the test period stays unseen for every model
Simulate all SL/TP combinations on that honest time window, fees included
Only a setup with real net profits gets switched to live

The AI has to prove it.

Built by tradersquant — TradingAI 12.0 ~Nexus, August 2026