# Model Configuration for Neuro-Symbolic Forecasting
# System 1: Neural Perception Layer (LSTM)

model:
  type: "LSTM"
  shared: true  # One model, all 50 stocks

lstm:
  input_size: 1
  hidden_size: 64
  num_layers: 2
  dropout: 0.2
  batch_first: true

output_layer:
  type: "Linear"
  output_size: 1

training:
  optimizer: "Adam"
  learning_rate: 0.001
  weight_decay: 0.0
  epochs: 15
  batch_size: 64
  early_stopping_patience: 3

data:
  sequence_length: 30  # 30-day look-back window
  forecast_horizon: 1   # Predict next 1 day
  train_split: 0.70
  val_split: 0.15
  test_split: 0.15
  normalization: "MinMaxScaler"
  scaler_range: [0, 1]

hardware:
  device: "cuda"
  multi_gpu: true
  gpus: 2
  mixed_precision: false

paths:
  model_save: "models/neural_checkpoint.pth"
  scaler_save: "models/scaler.pkl"
  best_model: "models/multi_stock_model.pth"

monitoring:
  log_frequency: 5  # Log every 5 epochs
  save_best: true
  best_metric: "val_loss"