normalization

Scaling in Stock Price Prediction with PyTorch

Neural Networks Are Sensitive to Feature Scale: Features with vastly different scales (e.g., stock prices in hundreds vs. RSI values between 0 and 100) cause uneven weight updates during training, which may slow convergence or cause instability.
Improves Gradient Descent Efficiency: Gradient descent converges faster and more reliably when input features are on a similar scale. Without scaling, some parameters may dominate updates, leading to poor local minima or slow training.
Prevents Numerical Instability: Large input values can cause exploding gradients or numerical overflow, especially in deep networks, while very small values can lead to vanishing gradients.
Enables Meaningful Regularization and Weight Initialization: Regularization techniques and common initializations assume input features have zero mean and unit variance, ensuring balanced learning.
Allows Model Generalization: Proper scaling helps the model learn from relative variations in features rather than absolute magnitudes, improving generalization to unseen data.

Common Scaling Techniques

Normalization (Min-Max Scaling): Rescales features to a fixed range, usually [0, 1].
X_scaled = (X - X_min) / (X_max - X_min)
Standardization (Z-score Scaling): Centers features by mean 0 and scales by standard deviation 1.
X_scaled = (X - μ) / σ

For financial time series, standardization is typically preferred due to outliers and varying ranges.

How to Effectively Scale Data in PyTorch

Assuming a dataset with multiple features for stock price prediction:

Step 1: Calculate Scaling Parameters on Training Data Only

import torch

	  # Example tensor: rows = samples, columns = features
	  train_features = torch.tensor([
	  [100.0, 0.7, 30.0],    # price, RSI, volume_change_pct
	  [102.0, 0.8, 25.0],
	  [98.0, 0.6, 40.0],
	  # ... more training samples
	  ])

	  # Compute mean and std only on training set for each feature
	  mean = train_features.mean(dim=0)
	  std = train_features.std(dim=0, unbiased=False)  # population std for consistency

Step 2: Apply Scaling to Train and Validation/Test Sets

def standardize(tensor, mean, std):
	  return (tensor - mean) / std

	  train_scaled = standardize(train_features, mean, std)

	  # When you get validation or test data:
	  val_features = torch.tensor([
	  [101.0, 0.75, 28.0],
	  # ...
	  ])
	  val_scaled = standardize(val_features, mean, std)

Important Notes

Do NOT compute scaling parameters on validation or test sets, only on training data to avoid data leakage.
If using minibatches, compute mean/std on the entire training set once, not per batch.
Save mean and std for inference to scale new data consistently.

After prediction, if your target was scaled, invert the scaling to recover original values:

def invert_standardize(scaled_tensor, mean, std):
	      return scaled_tensor * std + mean

Optionally, use torchvision.transforms.Normalize or sklearn.preprocessing.StandardScaler to automate scaling.

freeradiantbunny.org