The authors apply a large-scale deep learning model (recurrent neural network with Long Short-term Memory units) to high-frequency financial stock data in order to forecast the next price direction (up or down).
Results:
Not only do they uncover that the prediction accuracy for the deep learning model substantially increases by up to 10% contrary to high-frequency VAR-based models, but they also observe how a "universal" deep learning model trained on a pooled dataset of 500 stocks is able to capture relations between order flow and price variations which are common to all stocks in the dataset.
The universal deep learning model is able to outperform the stock-specific counterpart and most importantly is able generalize to stocks not included in the training set: if the model was trained on data for $\{1,\ldots,N\}$ stocks, then it can maintain the out-of-sample forecast accuracy for stock $N+1$. In conclusion, the model might viably predict the price direction on illiquid stocks where econometric models might falter due to lack of data.
The authors neatly summarizes the main findings of their deep learning model:
Nonlinearity: Data-driven models trained using
deep learning substantially outperform linear models in terms of forecasting accuracy (Section 3.1).
Universality: The model uncovers universal features that are common across all stocks (Section
3.2). These features generalize well: they are also
observed to hold for stocks which are not part of
the training sample.
Stationarity: The model performance in terms
of price forecasting accuracy is remarkably stable across time, even a year out of sample. This
shows evidence for the existence of a stationary
relationship between order flow and price changes
(Section 3.3), which is stable over long time
periods.
Path-dependence and long-range dependence: Inclusion of price and order flow history is shown to substantially increase the forecast accuracy. This provides evidence that price dynamics depend not only on the current or recent state of the limit order book but on its history, possibly over long time scales (Section 3.4).
It should be noted that the paper only focuses on the predictive accuracy of the deep learning model and do not determine whether this predictability leads to profitable trading strategies. Implementing transaction costs and accounting for latency in the high-frequency data would be out of the scope of the paper.
Additional Info: There is a Quant Stack Exchange question asking about appropriate machine learning models for financial time-series forecasting. Here, the highest voted answer has a good list of recent papers, including the paper provided in @MiLuk's answer.
