Event-Driven Sentiment Signals

date,event_type,score,source,summary
2024-01-15,earnings,0.8,read_url,Q4 revenue beat expectations by 30%
2024-01-20,macro,-0.5,read_url,Central bank raised rates by 25bp
2024-02-01,policy,0.3,read_url,New-energy subsidies extended
2024-02-10,sentiment,-0.7,read_url,Bearish sentiment surged on social media
2024-03-05,insider,0.4,read_url,CEO bought 5 million shares

import numpy as np
import pandas as pd


def compute_event_signal(event_df: pd.DataFrame, dates: pd.DatetimeIndex,
                         decay_lambda: float = 0.1,
                         min_score_threshold: float = 0.2,
                         event_lookback: int = 30) -> pd.Series:
    """Compute an event-driven signal with time decay.

    Args:
        event_df: DataFrame loaded from the event CSV, with date/event_type/score/source/summary columns.
        dates: Backtest date sequence (DatetimeIndex).
        decay_lambda: Decay coefficient. Higher values decay faster. Default 0.1 (decays to ~37% in about 10 days).
        min_score_threshold: Minimum score threshold. Events with |score| below this value are ignored.
        event_lookback: Event lookback window in days. Events older than this are excluded.

    Returns:
        Event signal Series aligned with dates, with value range [-1.0, 1.0].
    """
    event_df = event_df[event_df["score"].abs() >= min_score_threshold].copy()
    event_df["date"] = pd.to_datetime(event_df["date"])

    signal = pd.Series(0.0, index=dates)

    for trade_date in dates:
        # Only consider events published on or before trade_date (avoid look-ahead)
        mask = (event_df["date"] <= trade_date) & \
               (event_df["date"] >= trade_date - pd.Timedelta(days=event_lookback))
        relevant = event_df[mask]

        if relevant.empty:
            continue

        days_since = (trade_date - relevant["date"]).dt.days.values
        scores = relevant["score"].values
        # Exponential decay: score * exp(-lambda * days)
        decayed = scores * np.exp(-decay_lambda * days_since)
        # Sum multiple events and clip to [-1, 1]
        signal[trade_date] = np.clip(decayed.sum(), -1.0, 1.0)

    return signal

def combine_signals(tech_signal: pd.Series, event_signal: pd.Series,
                    alpha: float = 0.6) -> pd.Series:
    """Combine technical and event signals with weights.

    Args:
        tech_signal: Technical signal, range [-1.0, 1.0].
        event_signal: Event-driven signal, range [-1.0, 1.0].
        alpha: Weight of the technical signal, default 0.6 (technical primary, event secondary).

    Returns:
        Combined signal, range [-1.0, 1.0].
    """
    combined = alpha * tech_signal + (1 - alpha) * event_signal
    return combined.clip(-1.0, 1.0)

You are a financial event analyst. Read the following news / announcement and score its impact on the stock price.

Scoring scale:
- 1.0: extremely bullish (for example, earnings far above expectations, major favorable policy)
- 0.5: moderately bullish (for example, earnings slightly above expectations, favorable industry news)
- 0.2: mildly bullish
- 0.0: neutral (no obvious impact)
- -0.2: mildly bearish
- -0.5: moderately bearish (for example, earnings below expectations, tighter industry regulation)
- -1.0: extremely bearish (for example, accounting fraud, major violations, black-swan event)

Score strictly on the scale above. Output one number only. Do not explain.

News content:
{news_content}

Score: