Structure
A lesson captures two context snapshots – one at entry, one at exit – alongside the trader’s optional reasoning and the trade outcome. The multidimensional state of the market, paired with the trader’s intent and the result. The diff between entry and exit is where the learning signal lives.140 Fields Across 15 Categories
| Category | Fields | What It Captures |
|---|---|---|
| Trade Details | 17 | Entry/exit params, sizing, leverage, P&L, duration, exit reason |
| Market Data | 15 | Price, volume, OI, funding, momentum, spread |
| Technical Analysis | 15 | RSI, MACD, Bollinger, EMA alignment, ATR, volume profile, support/resistance, regime |
| Order Flow & Microstructure | 11 | Retail/whale volume delta, orderbook depth imbalance, SL/TP clusters, cross-exchange OI, funding percentile |
| Sentiment & Narrative | 9 | Social attention share, narrative momentum, mention velocity, sentiment shift, influencer signal |
| Positioning Intelligence | 9 | Cohort bias, divergence, exposure ratio, liquidation risk concentration, whale activity, trader count |
| Risk & Correlation | 9 | BTC correlation, portfolio directional alignment, liquidation clusters, position count, stress testing |
| On-Chain Intelligence | 8 | Institutional-scale perp flows, exchange netflows, new wallet activity, large-holder divergence |
| Macro Context | 8 | Fear & Greed, BTC price, BTC EMA state, BTC dominance, BTC funding, ETH/BTC ratio |
| Entity Intelligence | 7 | Labeled wallet flows, bridge volume, DeFi leverage exposure, counterparty distribution |
| Trader Context | 7 | Opening thesis, conviction level, time horizon, trade trigger, closing reflection |
| Trader Stats | 7 | Win/loss streak, recent PnL, trade frequency, win rate, avg hold time |
| Agent Metadata | 7 | Confidence score, thesis tag, conviction factors, flagged risks, swarm bias, stop reasoning |
| Execution Quality | 6 | Slippage, fill latency, order type, fill rate, session, day of week |
| Counterfactual Context | 5 | Top mover during hold, best alternative setup at entry, opportunity cost ranking |
Why This Structure
The lesson schema reflects established research traditions applied to a novel domain.- Live-capital data addresses backtest bias. Academic research has shown that backtested strategies systematically overstate real-world performance due to selection bias and execution frictions (Bailey & López de Prado, 2014). Engram captures lessons from live trades with real capital, producing training signal that avoids the selection bias and execution-frictionless assumptions that undermine backtest-based research.
- Multimodal capture reflects research on signal combination. The mathematical foundation for combining multiple independent signals is well-established (Grinold, 1989), and recent empirical studies confirm that multimodal fusion outperforms single-source approaches in financial prediction tasks. Engram’s 15 categories span market structure, positioning, microstructure, sentiment, and trader psychology, designed to maximize signal diversity.
- Reasoning paired with outcome extends modern reward modeling research. Work on process supervision (Lightman et al., 2023) has shown that training on reasoning alongside outcomes produces more reliable models than outcome-only training. The resulting state-action-reasoning-outcome data structure is compatible with offline reinforcement learning paradigms including Decision Transformer (Chen et al., 2021), behavioral cloning, and preference-based methods.
- Feature-level generalization across asset classes and rare events. Classical quantitative trading requires statistical regularity within narrow asset classes. Engram’s 140-field decomposition captures the properties of market events rather than the events themselves, enabling learning across crypto perps, tokenized equities, commodities, and prediction markets — asset classes that will increasingly trade side-by-side on Hyperliquid as HIP-3 and HIP-4 expand the venue’s scope.