Great question—this is a super common point of confusion when getting started with time-series stock prediction models, and the right approach boils down to what your neural network is designed to do. Let’s break this down clearly:

1. Default Approach: Normalize Using a Single Stock’s Historical Data #

If your goal is to predict the future price movement of an individual stock (e.g., "Will StockA’s price go up tomorrow based on its own past prices/volume?"), this is the way to go. Here’s why:

• Stock prices are relative to their own history: A $5 jump for a $10 stock is a 50% gain, while a $5 jump for a $100 stock is only 5%. Normalizing against the stock’s own past data lets your model learn these meaningful relative changes, instead of getting distracted by raw price differences between stocks.
• Avoids data leakage: In real-world trading, you can’t use other stocks’ current day prices to normalize your target stock’s data before making a prediction (you don’t have that data yet!). Using historical data only ensures your model trains on information that would actually be available at prediction time.
• Preserves stock-specific patterns: Each stock has its own volatility, trend, and volume behavior. Normalizing per stock keeps these unique patterns intact, which is critical for accurate single-stock forecasting.

Example Workflow #

For a single stock’s closing price, you’d typically use either:

• Min-Max Scaling (to squeeze values into a 0-1 range):

  from sklearn.preprocessing import MinMaxScaler
  import pandas as pd
  
  # Assume stock_df contains StockA's historical 'close' and 'volume' data
  scaler_price = MinMaxScaler(feature_range=(0, 1))
  scaler_volume = MinMaxScaler(feature_range=(0, 1))
  
  # Fit ONLY on training data to avoid leakage!
  stock_df['norm_close'] = scaler_price.fit_transform(stock_df[['close']])
  stock_df['norm_volume'] = scaler_volume.fit_transform(stock_df[['volume']])
  

• Z-Score Normalization (to center around historical mean, scaled by standard deviation):

  # Calculate mean/std from training data only
  train_mean = stock_df['close'].iloc[:train_size].mean()
  train_std = stock_df['close'].iloc[:train_size].std()
  
  stock_df['zscore_close'] = (stock_df['close'] - train_mean) / train_std
  

2. When to Use Daily Cross-Stock Normalization #

Only use this if your model’s goal is cross-stock comparison, like predicting which stock will outperform others on a given day (e.g., "Which of StockA/StockB/StockC has the highest probability of rising tomorrow?"). In this case:

• You need to put all stocks on the same scale to let the model compare their relative performance. For example, normalizing all stocks’ daily prices against the day’s market average or using a cross-stock Z-score.
• Critical note: You must still avoid data leakage! Never use the current day’s cross-stock stats to normalize training data. Instead, use rolling stats from prior days (e.g., normalize each day’s prices against the average of the past 30 days across all stocks).

Quick Summary #

Always remember: Fit your normalization scalers only on training data, then transform both training and test data with those pre-fit scalers. This is the #1 way to avoid accidental data leakage that makes your model look great in testing but fail in real trading.

内容的提问来源于stack exchange，提问作者BigBadMe