Signal Generation & Sentiment Mapping

This page describes the implementation of signal generation within the feb_2026_strategy. It focuses on how the system translates high-level LLM forecasts into actionable trade signals using a graph-based data resolution pattern, sentiment mapping, and strict filtering rules.

Sentiment to Position Mapping

The core of the strategy is the translation of natural language sentiment labels produced by the LLM into discrete trading positions. This is governed by the POSITION_LABEL_MAP constant.

LLM Sentiment	Trade Position	Strategy Action
`bullish`	`long`	Open Buy position
`bearish`	`short`	Open Sell position
`neutral`	`wait`	No action / Ignore
`sideways`	`wait`	No action / Ignore

This mapping ensures that only clear directional biases result in market exposure. The wait status acts as a safety buffer for ambiguous market conditions.

Graph-Based Data Resolution

The strategy utilizes a graph-based pattern from @backtest-kit/graph to manage data dependencies. This separates the fetching of the forecast from the logic that determines the position.

forecastSource: A sourceNode that wraps the forecast function. It includes a Cache.file layer with a 1-day interval to prevent redundant LLM calls during backtesting.
positionOutput: An outputNode that depends on forecastSource. It applies the POSITION_LABEL_MAP to the sentiment returned by the LLM.

Signal Generation Pipeline

The getSignal function is the entry point for the strategy logic, registered via addStrategySchema. It follows a sequential validation pipeline before emitting a signal.

1. Temporal Filtering (Cooldown)

To prevent over-trading on the same news cycle, the strategy implements a NEWS_WINDOW of 24 hours (1440 minutes).

The function getMinutesSinceLatestSignalCreated checks the time elapsed since the last signal for the specific symbol.
If the elapsed time is less than NEWS_WINDOW, the strategy returns null, effectively enforcing a cooldown period.

2. Forecast Resolution & Confidence Filtering

The strategy resolves the graph nodes to obtain the latest AI insights.

Confidence Check: If forecast.confidence is marked as not_reliable, the signal is discarded.
Sentiment Check: Explicit checks for neutral or sideways sentiments provide a secondary layer of protection against non-directional signals.

3. Signal Construction

If all filters pass, a signal object is returned.

ID: Generated by combining the forecast ID with a random string to ensure uniqueness.
Moonbag Pattern: Uses Position.moonbag to set the position side and a hard stop-loss (defined by HARD_STOP at 3.0%).
Metadata: A formatted Markdown note is attached, containing the LLM's reasoning and a deep link to the news dump for auditability.

Data Flow Diagram: Forecast to Signal

The following diagram illustrates the transition from the LLM's natural language output to the structured signal used by the execution engine.

Title: Signal Generation Logic Flow Mermaid Diagram

Sentiment Flip Detection

The strategy does not just open positions; it actively monitors them for sentiment reversals via listenActivePing. This creates a dynamic link between the LLM's evolving view of the market and open trades.

Continuous Polling: Every "ping" (tick) of the active position triggers a re-resolution of forecastSource and positionOutput.
Comparison: The current LLM-suggested position is compared against the actual data.position of the open trade.
Flip Execution: If the sentiment has changed (e.g., from long to short or wait), and the new forecast is reliable, the strategy calls commitClosePending to exit the position immediately.

Signal State Mapping Diagram

Title: Code Entities and State Transitions Mermaid Diagram