fin-glassbox

Hyperparameter Configuration

Document Purpose

This document contains all hyperparameter configurations for every model in the Explainable Multimodal Neural Framework for Financial Risk Management. These values are based on:

Version: 1.0
Status: Ready for Implementation

Table of Contents

  1. Global Training Configuration
  2. Encoder Layer
  3. Analyst Layer
  4. Risk Engine
  5. Fusion Layer
  6. Hyperparameter Search Spaces
  7. Learning Rate Schedules
  8. Regularization Summary
  9. YAML Configuration File

1. Global Training Configuration

Parameter Value Description
device "cuda" if available else "cpu" Compute device
seed 42 Random seed for reproducibility
dtype float32 Default tensor type
num_workers 4 DataLoader workers
pin_memory True Faster GPU transfer
checkpoint_dir "./checkpoints/" Model save location
log_dir "./logs/" TensorBoard logs
mixed_precision True FP16 training for speed

Training Chunks (Chronological)

training_chunks:
  chunk_1:
    train_years: [2000, 2001, 2002, 2003, 2004]
    val_years: [2005]
    test_years: [2006]
    description: "Dot-com recovery period"
  
  chunk_2:
    train_years: [2007, 2008, 2009, 2010, 2011, 2012, 2013, 2014]
    val_years: [2015]
    test_years: [2016]
    description: "Financial crisis + recovery"
  
  chunk_3:
    train_years: [2017, 2018, 2019, 2020, 2021, 2022]
    val_years: [2023]
    test_years: [2024]
    description: "COVID + bull market"

2. Encoder Layer

2A. Shared Temporal Attention Encoder

Parameter Value Search Range (TPE) Notes
d_model 128 [64, 128, 256] Embedding dimension
n_layers 4 [2, 3, 4, 5, 6] Transformer layers
n_heads 4 [2, 4, 8] Attention heads
d_ff 512 4 × d_model Feed-forward dimension
dropout 0.1 [0.05, 0.1, 0.15, 0.2] General dropout
attention_dropout 0.1 [0.05, 0.1, 0.15] Attention-specific dropout
activation "gelu" - GELU (standard for transformers)
max_seq_len 90 - Maximum lookback days
batch_size 32 [16, 32, 64] Per GPU
epochs 100 [50, 75, 100, 150] With early stopping
learning_rate 1e-4 loguniform(5e-5, 5e-4) Peak LR after warmup
weight_decay 1e-5 loguniform(5e-6, 5e-5) L2 regularization
warmup_steps 4000 [2000, 4000, 6000] Linear warmup
lr_schedule "cosine" - Cosine decay with warmup
gradient_clip 1.0 - Clip gradient norm
label_smoothing 0.05 [0.0, 0.05, 0.1] Soften labels
early_stop_patience 20 [10, 15, 20, 25] Epochs without improvement
optimizer "AdamW" - Adam with decoupled weight decay

TPE Trials: 50-100

2B. FinBERT Financial Text Encoder

Parameter Value Search Range Notes
base_model "ProsusAI/finbert" - Pre-trained FinBERT
max_length 512 - Max tokens per document
projection_dim 256 [128, 256, 384] Output embedding size
freeze_base False - Fine-tune entire model
batch_size 16 [8, 16, 32] Small due to memory
epochs_per_chunk 3 [2, 3, 4, 5] Fine-tuning epochs
learning_rate 2e-5 loguniform(1e-5, 5e-5) Standard BERT fine-tune LR
weight_decay 0.01 loguniform(0.001, 0.1) L2 regularization
warmup_proportion 0.1 [0.05, 0.1, 0.15] % of steps for warmup
gradient_clip 1.0 - Clip gradient norm
dropout 0.1 [0.05, 0.1, 0.15] Classifier dropout
early_stop_patience 5 [3, 5, 7] Epochs without improvement
optimizer "AdamW" - Adam with decoupled weight decay
scheduler "linear" - Linear decay with warmup

TPE Trials: 20-30

3. Analyst Layer

3A. Technical Analyst (BiLSTM)

Parameter Value Search Range (TPE) Notes
input_dim 128 - Temporal embedding size
hidden_dim 64 [32, 64, 128] LSTM hidden size
num_layers 1 [1, 2] BiLSTM layers
bidirectional True - Use both directions
dropout 0.3 [0.2, 0.3, 0.4] LSTM dropout
use_attention_pooling True - Weighted sequence pooling
output_dim 3 - Trend, momentum, timing
batch_size 64 [32, 64, 128] Per GPU
epochs 50 [30, 50, 75] With early stopping
learning_rate 1e-3 loguniform(5e-4, 5e-3) Adam LR
weight_decay 1e-4 loguniform(5e-5, 5e-4) L2 regularization
gradient_clip 1.0 - Clip gradient norm
early_stop_patience 20 [15, 20, 25] Epochs without improvement
optimizer "Adam" - Adam optimizer
scheduler "reduce_on_plateau" - Reduce LR when loss stalls

TPE Trials: 30-50

3B. Sentiment Analyst (MLP)

Parameter Value Search Range Notes
input_dim 256 - Text embedding size
hidden_dims [128, 64] [[128], [128, 64], [256, 128, 64]] Layer sizes
output_dim 2 - Polarity, confidence
dropout 0.2 [0.1, 0.2, 0.3] Regularization
activation "relu" - ReLU activation
batch_norm True - Batch normalization
batch_size 128 [64, 128, 256] Per GPU
epochs 30 [20, 30, 50] With early stopping
learning_rate 1e-3 loguniform(5e-4, 5e-3) Adam LR
weight_decay 1e-5 loguniform(5e-6, 5e-5) L2 regularization
label_smoothing 0.05 [0.0, 0.05, 0.1] Soften labels
early_stop_patience 15 [10, 15, 20] Epochs without improvement
optimizer "Adam" - Adam optimizer

Grid Search: 9 combinations → TPE Fine-tuning: 20 trials

3C. News Analyst (Multi-Head Attention Pooling)

Parameter Value Search Range (TPE) Notes
input_dim 256 - Per-document embedding size
num_heads 4 [2, 4, 8] Attention heads
head_dim 64 - Dimension per head
dropout 0.1 [0.05, 0.1, 0.15] General dropout
attention_dropout 0.1 [0.05, 0.1, 0.15] Attention dropout
output_dim 2 - Impact score, relevance
batch_size 64 [32, 64, 128] Per GPU (documents vary)
epochs 40 [30, 40, 50] With early stopping
learning_rate 1e-3 loguniform(5e-4, 5e-3) Adam LR
weight_decay 1e-5 loguniform(5e-6, 5e-5) L2 regularization
gradient_clip 1.0 - Clip gradient norm
early_stop_patience 15 [10, 15, 20] Epochs without improvement
optimizer "Adam" - Adam optimizer

TPE Trials: 20-30

4. Risk Engine

4A. Volatility Estimation (GARCH + MLP Hybrid)

GARCH Component

Parameter Value Notes
p 1 ARCH order
q 1 GARCH order
dist "normal" Error distribution
rolling_window 252 1 year of trading days
update_frequency "daily" Recalculate daily

MLP Component

Parameter Value Search Range (TPE) Notes
input_dim 128 - Temporal embedding size
hidden_dims [64] [[32], [64], [64, 32]] Hidden layer sizes
output_dim 4 - Vol_10d, Vol_30d, Regime, Confidence
dropout 0.2 [0.1, 0.2, 0.3] Regularization
activation "relu" - ReLU activation
batch_size 128 [64, 128, 256] Per GPU
epochs 40 [30, 40, 50] With early stopping
learning_rate 1e-3 loguniform(5e-4, 5e-3) Adam LR
weight_decay 1e-5 loguniform(5e-6, 5e-5) L2 regularization
early_stop_patience 15 [10, 15, 20] Epochs without improvement
optimizer "Adam" - Adam optimizer

TPE Trials: 30-50

4B. Drawdown Risk (BiLSTM Dual Horizon)

Parameter Value Search Range (TPE) Notes
input_dim 128 - Temporal embedding size
hidden_dim 64 [32, 64, 128] LSTM hidden size
num_layers 1 [1, 2] BiLSTM layers
bidirectional True - Use both directions
dropout 0.3 [0.2, 0.3, 0.4] LSTM dropout
output_dim_10d 3 - Prob, depth, recovery (10-day)
output_dim_30d 3 - Prob, depth, recovery (30-day)
batch_size 64 [32, 64, 128] Per GPU
epochs 50 [30, 50, 75] With early stopping
learning_rate 1e-3 loguniform(5e-4, 5e-3) Adam LR
weight_decay 1e-4 loguniform(5e-5, 5e-4) L2 regularization
gradient_clip 1.0 - Clip gradient norm
early_stop_patience 20 [15, 20, 25] Epochs without improvement
optimizer "Adam" - Adam optimizer
scheduler "reduce_on_plateau" - Reduce LR when loss stalls

TPE Trials: 30-50

4C. VaR & CVaR (Non-parametric)

No hyperparameters — statistical calculations only.

Parameter Value Notes
rolling_window 504 2 years of trading days
confidence_levels [0.95, 0.99] VaR thresholds
update_frequency "daily" Recalculate daily
method "historical" Empirical distribution

4D. GNN Contagion Risk (StemGNN)

Based on successful baseline reproduction results.

Parameter Value Search Range (TPE) Notes
num_nodes 4428 - Number of stocks
window_size 30 [15, 30, 60] Days of lookback
horizon 1 - Predict 1 step (contagion, not price)
multi_layer 13 [5, 8, 13, 20] StemGNN blocks
embed_size 32 [16, 32, 64] Node embedding dimension
hidden_size 64 [32, 64, 128] Hidden dimension
learning_rate 0.01 loguniform(0.001, 0.05) RMSprop LR
exponential_decay_step 13 [5, 8, 13] LR decay step
decay_rate 0.5 [0.3, 0.5, 0.7, 0.9] LR decay multiplier
dropout_rate 0.75 [0.5, 0.6, 0.75, 0.8] Regularization
batch_size 32 [16, 32, 64] Per GPU
epochs 100 [50, 75, 100] With early stopping
optimizer "RMSprop" - From baseline
norm_method "z_score" - Input normalization
early_stop_patience 20 [15, 20, 25] Epochs without improvement
gradient_clip 1.0 - Clip gradient norm
train_length 7 - Years for training split
valid_length 2 - Years for validation split
test_length 1 - Years for test split
leakyrelu_rate 0.2 - LeakyReLU slope
cheb_k 3 - Chebyshev polynomial order
top_k_edges 66 [20, 44, 66, 100] Edges per node (√4428 ≈ 66)

TPE Trials: 50-100

4E. Liquidity Risk (Rule-based)

No hyperparameters — rule-based thresholds only.

Parameter Value Notes
min_volume_percentile 20 Below this = illiquid
max_spread_pct 0.5 Above this = high slippage
market_cap_tiers [10e9, 2e9] Large (>$10B), Mid, Small (<$2B)
days_to_liquidate_threshold 5 >5 days = liquidity warning
update_frequency "daily" Recalculate daily

4F. Regime Detection (MTGNN Graph Builder + Classifier)

Graph Builder Component

Parameter Value Search Range Notes
num_nodes 4428 - Number of stocks
node_embedding_dim 64 [32, 64, 128] MTGNN embedding size
top_k 66 [20, 44, 66, 100] Edges per node
cheb_k 3 [2, 3, 5] Chebyshev order
input_dim 384 - 128 temporal + 256 text

Classifier Component

Parameter Value Search Range (TPE) Notes
input_features 5 - Density, modularity, avg_degree, clustering_coef, transitivity
hidden_dims [32] [[16], [32], [32, 16]] Hidden layer sizes
output_classes 4 - calm, volatile, crisis, rotation
dropout 0.2 [0.1, 0.2, 0.3] Regularization
activation "relu" - ReLU activation
batch_size 256 [128, 256, 512] Per GPU
epochs 30 [20, 30, 40] With early stopping
learning_rate 1e-3 loguniform(5e-4, 5e-3) Adam LR
weight_decay 1e-5 loguniform(5e-6, 5e-5) L2 regularization
early_stop_patience 10 [5, 10, 15] Epochs without improvement
optimizer "Adam" - Adam optimizer

Training Frequency: Weekly (graph building only)
TPE Trials: 20-30 (classifier only)

4G. Position Sizing Engine (Rule-based, User-adjustable)

No hyperparameters — user-configurable weights.

Parameter Default Value User Adjustable Notes
weight_volatility 0.20 Yes Weight for volatility risk
weight_drawdown 0.15 Yes Weight for drawdown risk
weight_var_cvar 0.15 Yes Combined VaR/CVaR weight
weight_contagion 0.25 Yes Weight for contagion risk
weight_liquidity 0.15 Yes Weight for liquidity risk
weight_regime 0.10 Yes Weight for regime risk
threshold_full 0.30 Yes Below this = 100% position
threshold_high 0.50 Yes Below this = 75% position
threshold_medium 0.70 Yes Below this = 50% position
threshold_low 0.85 Yes Below this = 25% position
veto_threshold 0.90 Yes Above this = no trade

5. Fusion Layer

Fusion Engine (MLP + Rules)

MLP Component (Layer 1)

Parameter Value Search Range (TPE) Notes
input_dim 13 - Qualitative(2) + Quantitative(2) + 9 module scores
hidden_dims [64, 32] [[32], [64, 32], [128, 64, 32]] Hidden layer sizes
output_dim 3 - Buy/Hold/Sell logits
dropout 0.2 [0.1, 0.2, 0.3] Regularization
activation "relu" - ReLU activation
batch_norm True - Batch normalization
batch_size 256 [128, 256, 512] Per GPU
epochs 50 [30, 50, 75] With early stopping
learning_rate 1e-3 loguniform(5e-4, 5e-3) Adam LR
weight_decay 1e-5 loguniform(5e-6, 5e-5) L2 regularization
label_smoothing 0.05 [0.0, 0.05, 0.1] Soften labels
early_stop_patience 15 [10, 15, 20] Epochs without improvement
optimizer "Adam" - Adam optimizer

Rule-based Component (Layer 2)

Rule Condition Action
liquidity_veto liquidity_score < 0.3 REJECT (no trade)
drawdown_cap drawdown_probability > 0.8 Cap size at 25%
contagion_veto contagion_score > 0.9 REJECT (no trade)
regime_override regime == “crisis” AND confidence > 0.7 Force SELL or HOLD

TPE Trials: 50-100

6. Hyperparameter Search Spaces

TPE (Bayesian Optimization) Configuration

tpe_config:
  algorithm: "tpe"  # Tree-structured Parzen Estimator
  n_initial_points: 20  # Random exploration before TPE
  n_trials:  # Model-specific (see below)
    temporal_encoder: 75
    finbert: 25
    fundamental_mlp: 30
    technical_analyst: 40
    sentiment_analyst: 20
    news_analyst: 25
    volatility_mlp: 40
    drawdown_bilstm: 40
    stemgnn: 75
    mtgnn_classifier: 25
    fusion_mlp: 75
  early_stop_trials: 20  # Stop if no improvement after 20 trials
  direction: "minimize"  # Minimize validation loss
  pruner: "median"  # Prune unpromising trials

Grid Search Configuration (for LightGBM/XGBoost)

grid_search_config:
  cv_folds: 5  # Chronological cross-validation
  scoring: "neg_mean_squared_error"
  n_jobs: -1  # Use all CPU cores
  verbose: 1

7. Learning Rate Schedules

Cosine Decay with Warmup (Temporal Encoder)

def cosine_decay_with_warmup(step, warmup_steps, total_steps, base_lr):
    if step < warmup_steps:
        # Linear warmup
        return base_lr * (step / warmup_steps)
    else:
        # Cosine decay
        progress = (step - warmup_steps) / (total_steps - warmup_steps)
        return base_lr * 0.5 * (1 + math.cos(math.pi * progress))

Reduce on Plateau (BiLSTM Models)

scheduler = ReduceLROnPlateau(
    optimizer,
    mode='min',
    factor=0.5,        # Reduce LR by half
    patience=10,       # Wait 10 epochs
    min_lr=1e-6        # Minimum LR
)

Linear Decay with Warmup (FinBERT)

def linear_decay_with_warmup(step, warmup_steps, total_steps, base_lr):
    if step < warmup_steps:
        return base_lr * (step / warmup_steps)
    else:
        return base_lr * (1 - (step - warmup_steps) / (total_steps - warmup_steps))

Exponential Decay (StemGNN)

def exponential_decay(epoch, base_lr, decay_step, decay_rate):
    return base_lr * (decay_rate ** (epoch // decay_step))

8. Regularization Summary

Model Dropout Attn Dropout Weight Decay Label Smooth Early Stop Grad Clip
Temporal Encoder 0.1 0.1 1e-5 0.05 20 1.0
FinBERT 0.1 0.1 0.01 - 5 1.0
Fundamental MLP 0.2 - 1e-5 - 15 -
Technical Analyst 0.3 - 1e-4 - 20 1.0
Sentiment Analyst 0.2 - 1e-5 0.05 15 1.0
News Analyst 0.1 0.1 1e-5 - 15 1.0
Volatility MLP 0.2 - 1e-5 - 15 1.0
Drawdown BiLSTM 0.3 - 1e-4 - 20 1.0
StemGNN 0.75 - 1e-5 - 20 1.0
MTGNN Classifier 0.2 - 1e-5 - 10 1.0
Fusion MLP 0.2 - 1e-5 0.05 15 1.0

XGBoost/LightGBM Regularization

Model L1 (reg_alpha) L2 (reg_lambda) Subsampling Early Stop
XGBoost (Fundamental Encoder) 0.1 1.0 0.7 50
LightGBM (Fundamental Analyst) 0.1 1.0 0.7 50

9. YAML Configuration File

# hyperparameters.yaml
# Complete hyperparameter configuration for all models

version: "1.0"
date: "2026-04-23"

global:
  device: "cuda"
  seed: 42
  dtype: "float32"
  num_workers: 4
  pin_memory: true
  mixed_precision: true
  checkpoint_dir: "./checkpoints/"
  log_dir: "./logs/"

training_chunks:
  chunk_1:
    train: [2000, 2001, 2002, 2003, 2004]
    val: [2005]
    test: [2006]
  chunk_2:
    train: [2007, 2008, 2009, 2010, 2011, 2012, 2013, 2014]
    val: [2015]
    test: [2016]
  chunk_3:
    train: [2017, 2018, 2019, 2020, 2021, 2022]
    val: [2023]
    test: [2024]

encoders:
  temporal:
    model: "transformer"
    d_model: 128
    n_layers: 4
    n_heads: 4
    d_ff: 512
    dropout: 0.1
    attention_dropout: 0.1
    activation: "gelu"
    max_seq_len: 90
    batch_size: 32
    epochs: 100
    learning_rate: 1.0e-4
    weight_decay: 1.0e-5
    warmup_steps: 4000
    lr_schedule: "cosine"
    gradient_clip: 1.0
    label_smoothing: 0.05
    early_stop_patience: 20
    optimizer: "AdamW"
    tpe_trials: 75

  finbert:
    base_model: "ProsusAI/finbert"
    max_length: 512
    projection_dim: 256
    freeze_base: false
    batch_size: 16
    epochs_per_chunk: 3
    learning_rate: 2.0e-5
    weight_decay: 0.01
    warmup_proportion: 0.1
    gradient_clip: 1.0
    dropout: 0.1
    early_stop_patience: 5
    optimizer: "AdamW"
    scheduler: "linear"
    tpe_trials: 25

  fundamental:
    xgboost:
      max_depth: 4
      learning_rate: 0.01
      n_estimators: 500
      subsample: 0.7
      colsample_bytree: 0.7
      reg_alpha: 0.1
      reg_lambda: 1.0
      min_child_weight: 5
      early_stopping_rounds: 50
      objective: "reg:squarederror"
      tree_method: "hist"
    mlp:
      input_dim: 70
      hidden_dims: [256]
      output_dim: 128
      dropout: 0.2
      activation: "relu"
      use_layer_norm: true
      batch_size: 128
      epochs: 50
      learning_rate: 1.0e-3
      weight_decay: 1.0e-5
      early_stop_patience: 15
      optimizer: "Adam"
      tpe_trials: 30

analysts:
  technical:
    model: "bilstm"
    input_dim: 128
    hidden_dim: 64
    num_layers: 1
    bidirectional: true
    dropout: 0.3
    use_attention_pooling: true
    output_dim: 3
    batch_size: 64
    epochs: 50
    learning_rate: 1.0e-3
    weight_decay: 1.0e-4
    gradient_clip: 1.0
    early_stop_patience: 20
    optimizer: "Adam"
    scheduler: "reduce_on_plateau"
    tpe_trials: 40

  sentiment:
    model: "mlp"
    input_dim: 256
    hidden_dims: [128, 64]
    output_dim: 2
    dropout: 0.2
    activation: "relu"
    batch_norm: true
    batch_size: 128
    epochs: 30
    learning_rate: 1.0e-3
    weight_decay: 1.0e-5
    label_smoothing: 0.05
    early_stop_patience: 15
    optimizer: "Adam"
    tpe_trials: 20

  news:
    model: "attention_pooling"
    input_dim: 256
    num_heads: 4
    head_dim: 64
    dropout: 0.1
    attention_dropout: 0.1
    output_dim: 2
    batch_size: 64
    epochs: 40
    learning_rate: 1.0e-3
    weight_decay: 1.0e-5
    gradient_clip: 1.0
    early_stop_patience: 15
    optimizer: "Adam"
    tpe_trials: 25

  fundamental_lgb:
    model: "lightgbm"
    input_dim: 128
    objective: "multiclass"
    num_class: 3
    boosting_type: "gbdt"
    num_leaves: 31
    learning_rate: 0.01
    n_estimators: 500
    subsample: 0.7
    colsample_bytree: 0.7
    reg_alpha: 0.1
    reg_lambda: 1.0
    min_child_samples: 20
    min_child_weight: 0.001
    early_stopping_rounds: 50
    verbosity: -1

risk:
  volatility:
    garch:
      p: 1
      q: 1
      dist: "normal"
      rolling_window: 252
      update_frequency: "daily"
    mlp:
      input_dim: 128
      hidden_dims: [64]
      output_dim: 4
      dropout: 0.2
      activation: "relu"
      batch_size: 128
      epochs: 40
      learning_rate: 1.0e-3
      weight_decay: 1.0e-5
      early_stop_patience: 15
      optimizer: "Adam"
      tpe_trials: 40

  drawdown:
    model: "bilstm"
    input_dim: 128
    hidden_dim: 64
    num_layers: 1
    bidirectional: true
    dropout: 0.3
    output_dim_10d: 3
    output_dim_30d: 3
    batch_size: 64
    epochs: 50
    learning_rate: 1.0e-3
    weight_decay: 1.0e-4
    gradient_clip: 1.0
    early_stop_patience: 20
    optimizer: "Adam"
    scheduler: "reduce_on_plateau"
    tpe_trials: 40

  var_cvar:
    rolling_window: 504
    confidence_levels: [0.95, 0.99]
    update_frequency: "daily"
    method: "historical"

  contagion:
    model: "stemgnn"
    num_nodes: 4428
    window_size: 30
    horizon: 1
    multi_layer: 13
    embed_size: 32
    hidden_size: 64
    learning_rate: 0.01
    exponential_decay_step: 13
    decay_rate: 0.5
    dropout_rate: 0.75
    batch_size: 32
    epochs: 100
    optimizer: "RMSprop"
    norm_method: "z_score"
    early_stop_patience: 20
    gradient_clip: 1.0
    train_length: 7
    valid_length: 2
    test_length: 1
    leakyrelu_rate: 0.2
    cheb_k: 3
    top_k_edges: 66
    tpe_trials: 75

  liquidity:
    min_volume_percentile: 20
    max_spread_pct: 0.5
    market_cap_tiers: [10.0e9, 2.0e9]
    days_to_liquidate_threshold: 5
    update_frequency: "daily"

  regime:
    graph_builder:
      num_nodes: 4428
      node_embedding_dim: 64
      top_k: 66
      cheb_k: 3
      input_dim: 384
    classifier:
      input_features: 5
      hidden_dims: [32]
      output_classes: 4
      dropout: 0.2
      activation: "relu"
      batch_size: 256
      epochs: 30
      learning_rate: 1.0e-3
      weight_decay: 1.0e-5
      early_stop_patience: 10
      optimizer: "Adam"
      tpe_trials: 25

  position_sizing:
    weights:
      volatility: 0.20
      drawdown: 0.15
      var_cvar: 0.15
      contagion: 0.25
      liquidity: 0.15
      regime: 0.10
    thresholds:
      full: 0.30
      high: 0.50
      medium: 0.70
      low: 0.85
      veto: 0.90

fusion:
  mlp:
    input_dim: 13
    hidden_dims: [64, 32]
    output_dim: 3
    dropout: 0.2
    activation: "relu"
    batch_norm: true
    batch_size: 256
    epochs: 50
    learning_rate: 1.0e-3
    weight_decay: 1.0e-5
    label_smoothing: 0.05
    early_stop_patience: 15
    optimizer: "Adam"
    tpe_trials: 75
  rules:
    liquidity_veto:
      condition: "liquidity_score < 0.3"
      action: "REJECT"
    drawdown_cap:
      condition: "drawdown_probability > 0.8"
      action: "CAP_SIZE_25%"
    contagion_veto:
      condition: "contagion_score > 0.9"
      action: "REJECT"
    regime_override:
      condition: "regime == 'crisis' AND confidence > 0.7"
      action: "FORCE_SELL_OR_HOLD"

tpe_config:
  algorithm: "tpe"
  n_initial_points: 20
  early_stop_trials: 20
  direction: "minimize"
  pruner: "median"

Document Version: 1.0
Status: Ready for Implementation
Next Step: Generate training scripts based on these configurations.

This site is open source. Improve this page.