Introducing Portfolio Transformer for Attention-Based Asset Allocation for after-work quants

Takeaway. The Portfolio Transformer (PT) skips predict-then-Markowitz: an encoder–decoder with attention, Time2Vec, and gated residual nets outputs signed weights (gross = 1) trained to maximise Sharpe after 2 bp costs.

Portfolio Transformer for Attention-Based Asset Allocation (Damian Kisiel, Denise Gorse, 2022; arXiv:2206.03246) is the after-work “train the allocation, not the forecast” paper.

Fit for learners: End-to-end Sharpe-optimizing transformer weights — skip predict-then-optimize and train allocation directly.

This Hub note is a plain-language intro. Not investment advice.

The problem

Two-step pipelines leak forecast error into weights. LSTMs struggle with long windows. Mean–variance needs a decent μ and Σ that we do not have. High turnover without a cost term in the loss looks brilliant in-sample and dies live.

The idea / core intuition

• Input: stacked daily returns of N assets over a window τ; Time2Vec encodes when.
• Encoder/decoder multi-head attention (causal mask in the decoder) learns cross-asset, long-range dependence.
• GRN/GLU can turn nonlinearity down when data are small or noisy.
• Output: signed-softmax of scores so shorts exist but sum of |w| = 1.
• Loss: Sharpe of cost-adjusted portfolio returns; expanding-window yearly refit in the paper.

Research on ETFs, commodities, and a large stock panel—not a product allocation.

Why after-work Python learners care

This is a reason to implement softmax on scores and a Sharpe loss before you touch BERT. Pair with Deep Momentum Networks: same “optimise the trading metric” religion, different architecture.

Tiny evening exercise: three toy scores s = [0.2, −0.5, 0.1]; compute signed-softmax weights and check sum(abs(w))==1. Learning only.

One-evening reading checklist

1. Why a separate return model injects error into Markowitz.
2. What Time2Vec adds that vanilla positional encodings miss.
3. How the output layer allows shorts without leverage.
4. Where transaction costs enter the loss.

Download PDF

Download PDF

Landing: arXiv abs

Important disclaimer

Not investment advice. This Hub note is educational only. Kisiel–Gorse is academic research—not a recommendation to replace your allocation with a Transformer.

What this is not

• Not CEF / subsidy framing
• Not Quant or Macro course CTA
• Not a robo-advisor pitch

Next step

Single course CTA: SoloQuant Academy’s Introduction to Python Programming—nested lists become return matrices; softmax is just exp and a sum.