CIPHER is currently in beta · Live testing in progress · Expect changes
CIPHERUpdates
CIPHER Changelog
New capabilities, improvements, and refinements to the CIPHER platform.
CIPHER-min-3
March 2026Major ReleaseNEW5,000 candles per pair (208 days / 7 months) — 10x more training data than min-2. Fetched in batches of 1,000 to work within Binance API limits.
NEWMulti-symbol rotation training — BTC, ETH, SOL, XRP rotate each episode. Symbol one-hot encoding in observation space lets the agent learn pair-specific patterns.
NEWRaw features fed directly to PPO (18 features) — removed LSTM latent from observation. Agent learns its own representations from RSI, MACD, BB, EMAs, returns, ATR.
NEWTest-split evaluation — python test_model.py --test-split runs on the held-out 15% across all 4 symbols. Current result: +3.6% return, 71.9% win rate, 32 trades.
NEWModel auto-versioning — saves a named snapshot every 500 episodes with win rate and return metadata.
NEWReward hacking detection — 10 consecutive zero-trade episodes or 3 zero-trade validations triggers automatic RL policy reset.
NEWTwo-database architecture — shared DB (knowledge, market data) + per-version training DB (episodes, trades). Safe to delete training DB without losing knowledge.
NEWLSTM internal validation — 80/20 train/val split with early stopping (patience=3). Weights only saved when validation loss improves.
FIXLSTM architecture overhauled — simplified for direction prediction (BCE). Retrain bug fixed that was corrupting weights.
FIXScaler data leakage fixed — MinMaxScaler now fitted on training split only. Validation and test data no longer leak future price ranges into training.
FIXLSTM padding fixed — zero padding instead of edge replication. Early episodes no longer see fabricated repeated prices.
FIXKnowledge query redundancy fixed — queried once per step instead of 2-3 times. Significant training speedup.
IMPROVEDRisk management tightened — 8 trades/episode (was 50), 5-candle cooldown (was 4), 100-candle max hold (was 30). Fewer but higher quality trades.
IMPROVEDStop loss widened to -3% and clamped — simulates a real stop-loss order. Prevents slippage beyond the stop.
IMPROVEDTake profit uses actual price move — if market gaps past 5%, the full move is captured.
IMPROVEDGamma reduced from 0.99 to 0.95 — agent focuses on shorter-term rewards, better suited for crypto volatility.
IMPROVEDRandom episode windows instead of sequential — prevents memorisation of fixed price sequences.
IMPROVEDValidation runs every 100 episodes across all 4 symbols with deterministic predictions. Weak validations trigger policy reset.
CIPHER-min-2
March 2026Major ReleaseNEWShort selling capability — CIPHER now profits from both rising and falling markets.
NEWDedicated Close Short action eliminates trading ambiguity and improves short-side precision.
NEWSmart exit rewards — CIPHER now voluntarily closes winning positions rather than waiting for time limits.
NEWLoss-cutting pressure — automatic escalating pressure to exit losing positions beyond 2% drawdown.
NEWData validation split — 70% training, 15% validation, 15% test. Prevents overfitting to historical data.
NEWVersion-isolated databases — each model version has its own isolated storage. No data contamination.
NEWVersion switcher in dashboard — toggle between current and archived versions without restarting.
IMPROVEDDashboard updates at the start of each episode rather than waiting for completion.
IMPROVEDKnowledge retrieval disabled during policy optimisation — cuts per-episode time in half.
IMPROVEDRSI values clamped to valid [0, 100] range for cleaner technical feature computation.
CIPHER-min-1
January 2026Stable ReleaseNEWFull knowledge store — 3,072 chunks from institutional trading literature, embedded and searchable.
NEWMulti-symbol training — BTC, ETH, SOL, XRP traded in rotation across sessions.
NEWLive Binance integration via ccxt — real market data, real fee simulation.
NEWReal-time dashboard with P&L charts, trade history, and per-pair stats.
NEWModel versioning — save, load, compare, and benchmark model versions at any point.
NEWREST API for external signal consumption.
IMPROVEDAnti-overtrading controls — maximum 2 trades per 10-candle window with escalating penalties.
IMPROVEDSession checkpoints every 100 episodes with automatic rollback capability.