Introducing Enhancing Time Series Momentum Strategies Using Deep Neural Networks for after-work quants

Takeaway. Deep Momentum Networks keep the time-series momentum (TSMOM) skeleton—trend signal × volatility scaling—but let a deep net learn the signal, often training with a Sharpe-like loss instead of next-return MSE.

Enhancing Time Series Momentum Strategies Using Deep Neural Networks (Bryan Lim, Stefan Zohren, Stephen Roberts, 2019; arXiv:1904.04912) is the after-work upgrade path from Moskowitz–Ooi–Pedersen-style TSMOM notebooks. Pair with Jegadeesh–Titman.

Fit for learners: Deep Momentum Networks with Sharpe loss + vol scaling — upgrade classical TSMOM in a weekend notebook.

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

The problem

Classical TSMOM uses a crude trend (sign of past excess return) and a vol target. The sign is jumpy; the mapping from history to position is hand-crafted. Supervised MSE on next returns also ignores that you care about risk-adjusted P&L, not point forecasts.

The idea / core intuition

• Keep vol scaling so positions shrink when risk spikes—an implementation detail that often dominates architecture.
• Replace or enrich the trend rule with LSTM / deep position outputs from lagged returns (and related features—see the paper).
• Train with objectives closer to portfolio Sharpe than to prediction MSE.
• Still compare against vanilla TSMOM; if the net cannot beat sign(past return) after costs, you learned something valuable.

Research upgrade—not a CTA sales sheet (the strategy family, not a course).

Why after-work Python learners care

You can implement the classical TSMOM loop in one evening of pandas, then swap the sign() for a tiny network. That is a healthier deep-learning project than predicting prices from scratch.

Tiny evening exercise: on one fake return series, compute 20-day momentum sign and 20-day vol; set toy position = sign / vol; print the strategy’s mean/vol of daily P&L vs buy-and-hold—learning only, not a futures book.

One-evening reading checklist

1. Write the classical TSMOM skeleton (trend × vol scale) in one line.
2. State what the deep net replaces (the hand-crafted trend).
3. Explain why a Sharpe-style loss differs from MSE.
4. List costs / turnover reasons a smoother net can still lose.

Download PDF

Download PDF

Landing: arXiv abs

Important disclaimer

Not investment advice. This Hub note is educational only. Lim–Zohren–Roberts is academic research—not a recommendation to trade time-series momentum with neural nets or real futures.

What this is not

• Not CEF / subsidy framing
• Not Quant or Macro course CTA
• Not a managed-futures product pitch

Next step

Single course CTA: SoloQuant Academy’s Introduction to Python Programming—loops over lists of numbers first; then a vol-scaled momentum loop is just homework.