model_utils.py

import os
import pandas as pd
import numpy as np
from sklearn.preprocessing import MinMaxScaler
from tensorflow.keras.models import load_model

#─── Data loading ──────────────────────────────────────────────────────────────

def load_and_merge_data():
    # Load stock data
    stock = pd.read_csv('scraped_combined_stock_data.csv')
    stock['Date'] = pd.to_datetime(stock['Date'])
    stock['Close'] = pd.to_numeric(stock['Close'], errors='coerce')
    stock = stock.dropna(subset=['Close'])

    # Load Reddit sentiment
    reddit = pd.read_csv('cleaned_reddit_data.csv')
    reddit['Date'] = pd.to_datetime(reddit['Date'])
    reddit = (reddit
              .groupby('Date')['sentiment']
              .mean()
              .reset_index()
              .rename(columns={'sentiment': 'reddit_sentiment'}))

    # Load Twitter sentiment
    twitter = pd.read_csv('cleaned_twitter_data.csv')
    twitter['Date'] = pd.to_datetime(twitter['Date'])
    twitter = (twitter
               .groupby('Date')['sentiment']
               .mean()
               .reset_index()
               .rename(columns={'sentiment': 'twitter_sentiment'}))

    # Drop tz info if present
    for df in (stock, reddit, twitter):
        df['Date'] = df['Date'].dt.tz_localize(None)

    # Merge all
    df = stock.merge(reddit, on='Date', how='left') \
              .merge(twitter, on='Date', how='left')
    df['reddit_sentiment'] = df['reddit_sentiment'].fillna(0)
    df['twitter_sentiment'] = df['twitter_sentiment'].fillna(0)
    return df

#─── Sequence creation ─────────────────────────────────────────────────────────

def create_sequences_multifeature(data: np.ndarray, time_steps: int = 60):
    X, y = [], []
    for i in range(len(data) - time_steps):
        X.append(data[i:i + time_steps])
        y.append(data[i + time_steps, 0])  # Close price as target
    return np.array(X), np.array(y)

#─── Preprocessing for a single ticker ─────────────────────────────────────────

def prepare_input(df, ticker: str, time_steps: int = 60):
    """
    Returns:
      last_sequence: np.array shaped (1, time_steps, 3)
      scaler: MinMaxScaler fitted on that ticker’s data
      error: None or error message
    """
    sub = df[df['Stock_Type'] == ticker]
    if len(sub) < time_steps:
        return None, None, f"Not enough data for {ticker}: {len(sub)} rows"

    features = sub[['Close', 'reddit_sentiment', 'twitter_sentiment']].values
    scaler = MinMaxScaler((0,1))
    scaled = scaler.fit_transform(features)

    # Grab last `time_steps`
    last_seq = scaled[-time_steps:].reshape(1, time_steps, scaled.shape[1])
    return last_seq, scaler, None

#─── Model loading + prediction ────────────────────────────────────────────────

def get_model_path(ticker: str, with_sentiment: bool = True):
    ext = 'with_sentiment' if with_sentiment else 'model'
    # you can adjust naming here if your files differ
    return f"lstm_{ticker}_{'model_with_sentiment' if with_sentiment else 'model'}.h5"

def predict_price(ticker: str, df, time_steps: int = 60):
    """
    Loads the model for `ticker`, preprocesses the last `time_steps` of data,
    runs a single-step forecast, and returns the denormalized price.
    """
    model_file = get_model_path(ticker, with_sentiment=True)
    if not os.path.exists(model_file):
        raise FileNotFoundError(f"Model file '{model_file}' not found.")

    seq, scaler, err = prepare_input(df, ticker, time_steps)
    if err:
        raise ValueError(err)

    model = load_model(model_file)
    pred_scaled = model.predict(seq)  # shape (1,1)
    # pad with zeros for the two sentiment dims so scaler inverse works
    padded = np.concatenate([pred_scaled, np.zeros((1,2))], axis=1)
    pred = scaler.inverse_transform(padded)[0,0]
    return float(pred)