Introducing Extending Deep Learning Models for Limit Order Books to Quantile Regression for after-work quants

Takeaway. DeepLOB-QR keeps the CNN-on-the-book backbone, then branches LSTMs so one network estimates several return quantiles for long and short sides at once—then a forecast combination of those quantiles can be more robust than a single point.

Extending Deep Learning Models for Limit Order Books to Quantile Regression (Zihao Zhang, Stefan Zohren, Stephen Roberts, 2019; arXiv:1906.04404) is the distributional sequel to DeepLOB.

Fit for learners: DeepLOB + quantiles for distributional mid-price moves — practical uncertainty layer for trading signals.

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

The problem

A mid-price direction classifier hides spread costs. Intraday long and short returns are not symmetric at the same stamp (you buy the ask and sell the bid). Fat tails mean a point forecast is not a risk picture. Training six separate nets on millions of LSE book events is wasteful if they share the same supply–demand features.

The idea / core intuition

• Define returns from first-level prices plus spreads, not mid-only; model long and short separately.
• Shared convolutional block on raw 10-level books; LSTM branches each train on a quantile loss (e.g. 0.25 / 0.5 / 0.75).
• Rearrange estimates if quantile crossing appears.
• Optionally combine quantile forecasts into a more stable centre (constrained weights summing to one).

Research template on 2017 LSE names—not a colocated bot.

Why after-work Python learners care

This is the cleanest reason to learn pinball / quantile loss after MSE: same tensors as DeepLOB, different heads. It also forces you to write long vs short PnL with spreads, which OHLCV notebooks skip.

Tiny evening exercise: for toy y = [−0.2, 0.1, 0.3], compute pinball loss at τ=0.25 vs τ=0.75 for a guess ŷ=0; write why the two losses are not symmetric. Learning only.

One-evening reading checklist

1. Why mid-price returns overstate what a taker actually earns.
2. Sketch the shared CNN + branched LSTM diagram.
3. Write the quantile (pinball) loss for one τ.
4. Name one reason combining quantiles can beat the median alone.

Download PDF

Download PDF

Landing: arXiv abs

Important disclaimer

Not investment advice. This Hub note is educational only. DeepLOB-QR is academic ML research—not a recommendation to trade quantile LOB nets with real capital.

What this is not

• Not CEF / subsidy framing
• Not Quant or Macro course CTA
• Not an HFT-as-a-service pitch

Next step

Single course CTA: SoloQuant Academy’s Introduction to Python Programming—lists and indexing first; quantile heads are just extra outputs on the same array.