AI Strategies

AI strategy overview

AI strategies use machine-learning models—LSTM, LightGBM, and Transformer—to learn price patterns from high-dimensional factor data and make data-driven trading decisions. Unlike traditional strategies, an AI strategy includes a dedicated model node in its workflow. Factors are calculated in the Data Processing node, passed to the model for training and prediction, and emitted as a prediction column (pred) for the strategy node.

All three examples below are production templates from Create Strategy → Strategy Templates, covering three use cases:

LSTM · Single-instrument timing
BTC · 1h · 96 factors
LightGBM · Long selection
Top 100 · 4h · Cross-sectional ranking
Transformer · Neutral hedge
Top 100 · 4h · Long-short neutral

💡 AI factor expressions are very long—33 technical factors plus multiple fundamental dimensions. Use View full configuration to expand the complete factor code, or Load to canvas to import the entire configuration into the strategy editor.

BTC timing — LSTM

LSTMSingle-instrument timingBTC · 1h · 96 factors

BTC Timing — LSTM

Uses an LSTM (long short-term memory network) for hourly BTC timing. Its 96 input factors consist of 49 technical factors, 22 common fundamentals, 17 BTC-specific factors, and 8 order-book factors. The model trains adaptively and predicts the direction of returns over the next six hours.

Core logic
49 technical factors
Multi-horizon momentum, volatility, tail risk, trend structure, price-volume behavior, liquidity, RSI/MACD/ADX, and interaction factors
22 common fundamentals
Funding rates, open interest, liquidations, options, exchange flows, on-chain activity, profit/loss cycles, sentiment, and ETF flows
17 BTC-specific factors
MVRV, SOPR, long/short-term holders, whale flows, illiquid supply, miner selling pressure, hash rate, and halving cycles
8 order-book factors
Spread anomaly, bid-ask imbalance and its 12-hour mean, normalized depth change, relative depth capacity, slope skew, impact asymmetry, and impact-cost anomaly. Shared with BTC/ETH LightGBM and ETH LSTM templates; historical order-book data is required
LSTM modeling
Adaptive configuration (configMode=adaptive) with a binary classification task; go long when pred > 0.01 and short when pred < -0.01
Workflow node configuration

Multifactor long — LightGBM

LightGBMLong selectionTop 100 · 4h · Cross-sectional ranking

Multifactor Long — LightGBM

Uses a LightGBM ranking model to rank instruments cross-sectionally within a Top 100 safety-screened universe. Inputs include 33 technical factors plus multi-instrument-compatible fundamental factors compiled for the 4h interval. The label is the return over the next six periods—about 24 hours at 4h resolution—with ranking conversion handled by the engine. The model ranks instruments by pred, buys the top five with equal weights, and rebalances every four hours.

Core logic
33 technical factors
Volatility, microstructure, price-volume behavior, momentum, trend, oscillators, statistics, patterns, channels, and buying pressure cover all major dimensions of market information
Multi-instrument fundamentals
Cross-instrument-compatible factors such as on-chain activity, exchange flows, and market-cap changes, compiled adaptively for the 4h interval
Unified return label
label = shift(close, -6) / shift(open, -1) - 1; the engine performs the label conversion required by the ranking task
LightGBM ranking task
The ranking task learns relative cross-sectional strength; a higher pred indicates a stronger expected return, and the top five are held long at equal weights
Workflow node configuration

Neutral hedge — Transformer

TransformerNeutral hedgeTop 100 · 4h · Long-short neutral

Neutral Hedge — Transformer

Uses a Transformer attention model to rank a Top 100 safety-screened universe cross-sectionally, buying the top five and shorting the bottom five for market-neutral exposure. It shares the factor system and universe of strategy #33 but uses a different model and position mode: the Transformer captures factor interactions and long-range temporal dependencies, while the neutral subtype automatically hedges beta.

Core logic
Transformer attention
Multi-head self-attention captures high-order interactions among the 33 technical factors and fundamental factors, learning nonlinear combinations
Unified return label
Shares the LightGBM strategy label: shift(close, -6) / shift(open, -1) - 1; the engine handles ranking conversion
Long-short neutral positions
Buy the five instruments with the highest pred and short the five with the lowest pred, allocating equal weights to hedge systematic market risk
Strict risk controls
Maximum 10% per instrument, five instruments on each side for at most ten holdings, and 80% total exposure to reduce concentration risk
Workflow node configuration
    BeeQuant - AI Quantitative Trading Platform