CIPHERWhat is CIPHER?
CIPHER is an autonomous AI trading model built to trade crypto markets 24/7 without human intervention. It watches four pairs on Binance, generates signals using a combination of deep learning and institutional knowledge retrieval, and executes trades in both directions — long and short.
CIPHER is currently in beta. The model is live and actively trading. After 90,000+ training episodes, the current win rate on held-out test data sits at 71.9% across all four pairs. The model is profitable after fees and continues to improve.
Current Status
CIPHER is in active beta testing. Live trading is running. Users on the early access whitelist are being onboarded in small batches. Expect updates, model version bumps, and performance improvements on a rolling basis.
Supported Pairs
CIPHER monitors and trades four pairs on Binance:
- BTC/USDT — Bitcoin
- ETH/USDT — Ethereum
- SOL/USDT — Solana
- XRP/USDT — XRP
Capital rotates between pairs based on signal confidence. CIPHER evaluates each pair independently and only deploys where the signal is clearest.
Version History
CIPHER has gone through three major versions. Each version improved win rate, reduced max drawdown, and generated higher cumulative returns over the same number of trading episodes.
Cumulative P&L ($k) — v1 vs v2 vs v3 · 90k episodes · $10k/episode
v3 cumulative: +$856kAgent learned to do nothing. Stuck at -$0.1k for 3,000+ episodes. Grade: D
Bug fixes: RSI clipped, 4-action space, train/val split. Modest improvement.
90k episodes, 4 symbols. Recovered from -$400k to +$856k. Test: 71.9% WR, +3.6% return.
v1 established the base architecture. v2 introduced short selling and improved risk controls. v3 (current) overhauled LSTM architecture, added multi-symbol rotation training, tightened risk management to 8 trades per episode, and widened stop loss to -3% with clamping — resulting in fewer but higher-quality trades and a 71.9% win rate on unseen data.
How It Learns
Unlike rule-based bots that follow fixed conditions, CIPHER actually learns. It trains on 5,000 candles (208 days) of history per pair, computes 18 technical features (RSI, MACD, Bollinger Bands, EMAs, multi-horizon returns, volume ratio, ATR, candle body), validates every 100 episodes against held-out data it has never seen, and checks a knowledge base of 3,072 insights extracted from institutional trading literature — then makes a call.
Most algorithmic systems stop at RSI, MACD, and Bollinger Bands. CIPHER uses all of those — but before every decision it also queries a curated library of trading insights. If the current setup matches patterns described in institutional research, that match shifts the confidence score.
Long & Short
CIPHER makes money in both directions. It opens long positions when it expects a rise, and short positions when it expects a drop. Each direction has its own dedicated entry and exit action — no ambiguity in what the signal means.
Fees & Realism
Every P&L number you see is after Binance's 0.1% fee. We don't show numbers that look good on paper but evaporate when you actually trade. CIPHER is trained to be profitable after transaction costs, not before.
Built By
CIPHER is designed and built by @itsalimanuel. All model versions, training runs, and architecture decisions are made by a single developer focused on building a genuinely useful trading system — not a demo.