Library / Machine learning in finance
Deep Learning on the Order Book, and Why It Does Not Apply Here
What does an order-book neural network predict that a simpler model misses?
The cited networks predict short-horizon price movements from event-level limit-order-book states and their recent history, with evidence of transfer across stocks. A daily OHLCV series lacks those inputs and event counts; compare any daily-bar model on its own data and horizon before inferring usefulness.
The paper studies order-book prediction; any claim about available app execution requires a separate code test.
Evidence map
| Aspect | Finding |
|---|---|
| What it is | The two results that made deep learning credible in microstructure, included in this pack mainly so that an agent can explain precisely why they do not transfer to an application that works from bars alone. |
| Key result / formula | Sirignano and Cont train a model to predict short-horizon price moves from the state and recent history of the limit order book, across a large set of United States equities. |
| Why it matters for backtesting | The sound use of this note is negative and specific. |
Key result / formula
Their central finding is universality: a single model trained on all stocks outperforms models trained on each stock individually, and it transfers to stocks absent from the training set, which indicates the relationship between order flow and price movement has a common form across instruments. They also find the mapping is nonlinear and that the history of the book carries information beyond its current state. Zhang, Zohren and Roberts build a convolutional architecture with recurrent layers for the same problem, taking raw book levels as input with no engineered features, and report that the learned features transfer to instruments absent from training. Both results depend on data of very high frequency and very large volume: millions of events per instrument, with the full depth of the book.
Why it matters for backtesting
These results rest on three things Stochastly does not have: the order book, event-level timestamps, and a sample size measured in millions of events per instrument. A user citing them as evidence that deep learning works in finance is citing a domain where data is abundant and the horizon is seconds, and transferring the conclusion to daily bars inverts each condition that made it work. What does transfer is the universality idea, and it is testable here in a weak form: a model trained across all of the user's instruments jointly, rather than one per instrument, uses the scarce data better and is less prone to fitting one series, which is the same argument that favours pooled evidence (see [Pooled Evidence on Correlated Units]).
Source
Sirignano & Cont, "Universal features of price formation in financial markets: perspectives from deep learning", Quantitative Finance 19(9), 2019, 1449-1459; Zhang, Zohren & Roberts, "DeepLOB: Deep Convolutional Neural Networks for Limit Order Books", IEEE Transactions on Signal Processing 67(11), 2019, 3001-3012. Primary source