Few Shot Learning Finance

Zero-Shot Learning:

• Model has NEVER seen the target asset during training
• Must transfer knowledge from different assets entirely
• Training set and test set disjoint: I_train ∩ I_test = ∅
• Example: Trading a new cryptocurrency using patterns learned from equities

# Few-shot setting
train_assets = ['SPY', 'GLD', 'TLT']  # 30 assets
test_assets = ['SPY', 'GLD', 'TLT']   # Same 30 assets, different time period

# Zero-shot setting
train_assets = ['SPY', 'GLD', 'TLT']  # 30 assets for training
test_assets = ['BTC', 'ETH', 'SOL']   # 20 different assets for testing

2. Episodic Learning

Train models the same way they'll be used at test time:

Traditional Training:

# Standard mini-batch training - all assets mixed together
for epoch in epochs:
    for batch in shuffle(all_data):
        loss = model(batch)
        optimizer.step()

Episodic Training:

# Episode-based training - mimics test-time usage
for episode in episodes:
    # Sample target sequence (what we want to predict)
    target_asset, target_time = sample_target()

    # Sample context set C (what we condition on)
    context_set = sample_contexts(
        assets=train_assets,
        exclude=(target_asset, target_time),  # Ensure causality
        size=C  # Number of context sequences
    )

    # Make prediction using context
    prediction = model(target=target, context=context_set)

    loss = criterion(prediction, true_value)
    optimizer.step()

Key Principles:

• Each episode = one prediction task
• Context set must be causal (occurred before target)
• Model learns to transfer patterns from context to target
• Trains on k-shot tasks to perform well on k-shot evaluation

3. Context Set Construction

Context set C contains sequences from other assets/regimes that inform the prediction.

Properties:

• Size: Typically 10-30 sequences
• Causality: All context must occur before target time
• Diversity: Include different assets and market conditions
• Quality: CPD segmentation improves performance 11.3% vs random

Construction Methods:

1. Random: Sample random sequences before target_time
2. Time-equivalent: Same time window as target, different assets
3. CPD-segmented: Use change-point detection for clean regime segments

See IMPLEMENTATION.md for code examples.

4. Meta-Learning Architecture

How It Works:

1. Encode context sequences → Learn patterns from similar situations
2. Encode target sequence → Understand current market state
3. Cross-attention → Target queries context for relevant patterns
4. Combine representations → Integrate transferred knowledge
5. Predict position → Generate trading signal

Key Insight: Cross-attention automatically identifies which context sequences are most similar to the target, weighting them higher in the final prediction.

See IMPLEMENTATION.md for implementation.

5. Transfer Learning Scenarios

1. Same Asset, Different Regime (Few-Shot)

• Target: SPY in 2020 (COVID crash)
• Context: SPY in 2008 (financial crisis), SPY in 2018 (correction)
• Transfer: Crisis response patterns

2. Different Assets, Similar Dynamics (Zero-Shot)

• Target: New cryptocurrency (BTC)
• Context: Gold, Silver, Crude Oil (commodities)
• Transfer: Trending behavior, volatility patterns

3. Cross-Asset Momentum Spillover

• Target: European equities (CAC40)
• Context: US equities (SPY), Asian equities (Nikkei)
• Transfer: Leading indicators, correlation structures

6. Training Objectives

Joint Loss Function:

L_joint = α * L_MLE + L_Sharpe