app/services/prediction.py

from typing import List
import numpy as np
from fastapi import HTTPException
from tensorflow.keras.models import load_model
from datetime import datetime, timedelta

from app.config import get_settings
from app.indicators import classify_signal, classify_volatility
from app.services.market_data import get_enriched_history, last_n_candles

settings = get_settings()

_model = None


def get_model():
    """
    Lazy-load model to avoid HF Space startup crashes
    """
    global _model
    if _model is None:
        try:
            _model = load_model(settings.model_path)
        except Exception as e:
            print(f"[Prediction] Failed to load model: {e}")
            raise HTTPException(
                status_code=500,
                detail="Prediction model not available",
            )
    return _model


def is_model_available() -> bool:
    """
    Used by health endpoint.
    Does NOT force model loading failure to crash the app.
    """
    try:
        get_model()
        return True
    except Exception:
        return False


# Normalization
def _normalize_window(window: np.ndarray):
    """
    window shape: (history_window, 5)
    [Close, RSI, MACD, Log_Volume, MA50]
    """
    start_p = window[0, 0]

    if start_p <= 0:
        raise HTTPException(
            status_code=500,
            detail="Invalid price normalization window",
        )

    vol_mean = np.mean(window[:, 3])
    if vol_mean <= 0:
        vol_mean = 1.0  # safety guard

    norm = np.zeros_like(window)
    norm[:, 0] = (window[:, 0] / start_p) - 1
    norm[:, 1] = window[:, 1] / 100.0
    norm[:, 2] = window[:, 2] / start_p
    norm[:, 3] = (window[:, 3] / vol_mean) - 1
    norm[:, 4] = (window[:, 4] / start_p) - 1

    return norm, start_p


# Prediction
def predict_path(ticker: str, days: int):
    model = get_model()

    if days > settings.max_forecast_days:
        days = settings.max_forecast_days

    df_raw, df_tech = get_enriched_history(ticker)

    if len(df_tech) < settings.history_window:
        raise HTTPException(
            status_code=400,
            detail=(
                f"Not enough clean data after indicators "
                f"(required={settings.history_window}, got={len(df_tech)})"
            ),
    )


    features = df_tech[
        ["Close", "RSI", "MACD", "Log_Volume", "MA50"]
    ].values

    last_window = features[-settings.history_window:]
    current_price = float(last_window[-1, 0])

    temp_window = last_window.copy()
    forecast_points: List[dict] = []

    for i in range(days):
        norm, start_p = _normalize_window(temp_window)

        pred_pct = model.predict(
            norm.reshape(1, settings.history_window, 5),
            verbose=0,
        )[0][0]

        # 🔒 Safety clamp
        pred_pct = float(np.clip(pred_pct, -0.2, 0.2))

        next_price = float(start_p * (1 + pred_pct))

        volatility = 0.015 * (i + 1)
        lower_bound = float(next_price * (1 - volatility))
        upper_bound = float(next_price * (1 + volatility))

        forecast_points.append(
            {
                "day": i + 1,
                "price": next_price,
                "lower": lower_bound,
                "upper": upper_bound,
            }
        )

        new_row = temp_window[-1].copy()
        new_row[0] = next_price
        temp_window = np.vstack([temp_window[1:], new_row])


    # Summary

    target_price = forecast_points[-1]["price"]
    predicted_change = ((target_price - current_price) / current_price) * 100

    vol_label = classify_volatility(df_raw)
    signal = classify_signal(predicted_change)

    rsi_latest = float(df_tech["RSI"].iloc[-1])
    macd_latest = float(df_tech["MACD"].iloc[-1])
    macd_label = "Bullish" if macd_latest >= 0 else "Bearish"

    historical = last_n_candles(df_raw, settings.history_window)

    prediction_payload = {
        "ticker": ticker.upper(),
        "currentPrice": current_price,
        "targetPrice": target_price,
        "predictedChange": predicted_change,
        "signal": signal,
        "rsi": rsi_latest,
        "macd": macd_label,
        "volatility": vol_label,
        "historicalPrices": historical,
        "predictedPrice": target_price,
        "forecastData": [
            {
                "day": p["day"],
                "price": p["price"],
                "changePct": ((p["price"] - current_price) / current_price) * 100,
            }
            for p in forecast_points
        ],
    }


    # Chart payload

    last_date = datetime.strptime(historical[-1]["date"], "%Y-%m-%d")

    chart_forecast_payload = {
        "ticker": ticker.upper(),
        "points": [
            {
                "date": historical[-1]["date"],
                "price": current_price,
                "lower": current_price,
                "upper": current_price,
            }
        ],
    }

    for i, p in enumerate(forecast_points, start=1):
        d = last_date + timedelta(days=i)
        chart_forecast_payload["points"].append(
            {
                "date": d.strftime("%Y-%m-%d"),
                "price": p["price"],
                "lower": p["lower"],
                "upper": p["upper"],
            }
        )

    return prediction_payload, chart_forecast_payload