AI large language model
Generate trading signals with a three-input prompt architecture and six mid-tier models with first-party web access.
LLM overview
The AI large language model node uses a three-input prompt architecture. It feeds factor data, market intelligence, and custom user instructions to the model in separate layers, then combines technical and fundamental information to generate a trading signal. It supports only six mid-tier models—DeepSeek, GLM, Kimi, Qwen, Claude Sonnet 5, and GPT-5.6 Terra—and obtains current information through each provider API's native web-access capabilities.
The LLM node outputs signed confidence in the range [-1.0, +1.0]. Its format is identical to LightGBM and LSTM model output and can connect directly to downstream trading nodes.
Use cases
Paper trading / Live trading
Analyze the latest factor values in real time for live-market trading workflows.
Testnet live trading
Uses the same logic as live trading and is intended for strategy validation.
Historical backtesting is not supported — model analysis depends on current market context, including news, sentiment, and on-chain data, which cannot be reconstructed reliably during a backtest. The output DataFrame contains only the latest timestep, so the backtest engine cannot replay it bar by bar.
Three-input architecture
Input 1: Factor interpretation (generated dynamically)
At runtime, factors are grouped and displayed as factor name + formula + value. When a formula is available it is included; otherwise only the name and value are shown. The model evaluates operator semantics, field meaning, and value strength together.
Input 2: Market intelligence (system-defined)
The system uses fixed retrieval dimensions and searches each asset separately. Dimensions already represented by upstream factors are skipped automatically, and only missing information such as events, public sentiment, and macro context is added. Every selectable model supports native web search through its provider API.
Input 3: Custom user instructions (editable)
Optionally add analysis preferences, risk-control rules, or external decision criteria, for example: “Prioritize unusual funding rates” or “Use the daily trend as the primary signal.”
Supported models
| Model | Description |
|---|---|
| deepseek-v4-pro | DeepSeek's latest Pro tier with native web search |
| glm-5-turbo | Zhipu GLM with native web search |
| kimi-k2.6 | Kimi with native web search |
| qwen3.7-plus | Qwen with native web search |
| claude-sonnet-5 | Anthropic Claude with native web search |
| gpt-5.6-terra | OpenAI mid-tier model with native web search |
The model-name prefix automatically routes requests to the corresponding API endpoint; there is no need to configure base_url manually.
Node configuration
| Setting | Default | Description |
|---|---|---|
| AI model | qwen3.7-plus | Select a model; the system routes the API endpoint automatically |
| API Token | — | API key for the selected model |
| Prediction column name | pred | Name of the prediction column in the output DataFrame |
| Data interval | 1h | Time granularity of the input factor data |
| Prediction periods | 4 | Forecast the next n periods; interval=1h and periods=4 forecasts four hours |
| Request timeout | 600s | Keep 300–600 seconds for web retrieval across multiple assets |
| Custom user instructions | Empty | Optional additional analysis preferences supplied as Input 3 |
Output format
The AI large language model node returns a standard DataFrame containing only rows from the latest timestep.
timestamp | symbol | factor columns... | pred 2026-02-28 | BTC_USDT | ... | +0.70 2026-02-28 | ETH_USDT | ... | -0.45
| pred value | Meaning |
|---|---|
| +0.70 | Bullish with 70% confidence strength |
| -0.45 | Bearish with 45% confidence strength |
| 0.00 | Neutral, wait, or prediction failed |
Runtime logs
Runtime emits structured logs in this sequence: startup configuration → factor loading and trimming → per-asset analysis → summary. The analysis text in each conclusion is the model's complete reasoning after combining factor data with market intelligence.
Analysis engine started > Qwen qwen3.7-plus | interval 1 hour | forecast window 4 hours Formula parsing: formulas matched for 3/5 columns; dropped 2 factors without formulas Loaded factors: 5 factors x 158 rows > 145 rows after cleaning (removed 13 NaN/Inf rows) Web search enabled -- retrieving the latest market developments in real time [1/3] Analyzing BTC_USDT... [OK] [1/3] BTC_USDT ▼ Bearish 60% (2.8s) -- RSI near oversold, bearish MACD crossover... [2/3] Analyzing ETH_USDT... [OK] [2/3] ETH_USDT ▲ Bullish 55% (2.6s) -- On-chain activity is recovering... Analysis complete | 2/2 succeeded | total time 5.7s