Source-linked AI summary
Universal features of price formation in financial markets: perspectives from Deep Learning
Justin Sirignano, Rama Cont
TL;DR
The paper examines whether price formation can be modeled as a universal, stationary relation rather than through asset-specific recent-data models. It trains a Deep Learning model on high-frequency order-book histories to predict next price moves, finding stable cross-stock and cross-time performance, including on unseen stocks. Longer histories improve forecasts, while the reported stationarity concerns the input-output relation rather than necessarily the inputs themselves.
Problem
The paper asks whether the price-formation relation is universal across assets and stable over time, despite common asset-specific modeling on recent data.
Method
A recurrent Deep Learning model uses current order-book states and nonlinear representations of past order-flow and price information to predict the next price move.
Results
The universal model is stable across stocks, sectors, and time, outperforms stock-specific models, generalizes to unseen stocks, and benefits from histories extending several hours.
Takeaways & Limitations
The findings support pooling financial data across stocks and using universal models for newly issued or data-limited securities.
Takeaways & Limitations
The reported stationarity applies to the relation between inputs and forecasts; the input variables themselves may remain non-stationary.
Abstract
from arXiv · showhide
Using a large-scale Deep Learning approach applied to a high-frequency database containing billions of electronic market quotes and transactions for US equities, we uncover nonparametric evidence for the existence of a universal and stationary price formation mechanism relating the dynamics of supply and demand for a stock, as revealed through the order book, to subsequent variations in its market price. We assess the model by testing its out-of-sample predictions for the direction of price moves given the history of price and order flow, across a wide range of stocks and time periods. The universal price formation model is shown to exhibit a remarkably stable out-of-sample prediction accuracy across time, for a wide range of stocks from different sectors. Interestingly, these results also hold for stocks which are not part of the training sample, showing that the relations captured by the model are universal and not asset-specific. The universal model --- trained on data from all stocks --- outperforms, in terms of out-of-sample prediction accuracy, asset-specific linear and nonlinear models trained on time series of any given stock, showing that the universal nature of price formation weighs in favour of pooling together financial data from various stocks, rather than designing asset- or sector-specific models as commonly done. Standard data normalizations based on volatility, price level or average spread, or partitioning the training data into sectors or categories such as large/small tick stocks, do not improve training results. On the other hand, inclusion of price and order flow history over many past observations is shown to improve forecasting performance, showing evidence of path-dependence in price dynamics.
1 Price formation: how market prices react to supply and demand
The paper asks whether price formation is universal and stationary across assets and time, challenging asset-specific models built on recent windows. Using Deep Learning, it finds evidence that pooling stocks captures stable, nonlinear relations between order flow, price history, and subsequent price changes.
- Research question: The central question is whether the price-formation map is independent of the specific asset and stable over time.Existing practice generally estimates each asset’s parameters from recent data because financial series are often treated as non-stationary.
- Motivation: Pooling data across assets and time could provide richer estimation if the relation between order flow and price fluctuations is universal and stationary.The paper motivates this possibility with the example that one asset’s severe order-flow imbalance could inform predictions for another asset.
- Approach: Deep Learning provides a nonparametric approach for estimating nonlinear relations between order flow and market-price fluctuations from large datasets.The approach uses multilayer neural networks trained by supervised learning.
- Findings: The universal model shows stable out-of-sample prediction accuracy across time, sectors, and stocks outside the training sample.This supports the claim that the learned relations are universal rather than asset-specific.
- Findings: The universal model outperforms stock-specific linear and nonlinear models in out-of-sample prediction accuracy.The result favors pooling financial data from various stocks over designing stock- or sector-specific models.
- Findings: Standard normalizations and sector or tick-size partitioning do not improve training results, whereas longer price and order-flow histories improve forecasting performance.The latter finding provides evidence of path-dependence in price dynamics.
2 A data-driven model of price formation via Deep Learning
The paper models next price moves from high-frequency limit-order-book histories using recurrent neural networks. Its architecture combines nonlinear processing with temporal memory and is trained on large pooled datasets, including comparisons with asset-specific models.
- Data: The dataset records orders, transactions, and cancellations for approximately 1000 NASDAQ stocks at high frequency.Electronic order books can update millions of times per day, producing very large datasets for modeling price formation.
- Data: The limit order book encodes visible supply and demand through outstanding buy and sell orders, their prices, and their available sizes.The best ask is the lowest sell order and the best bid is the highest buy order; their difference is the spread.
- Model architecture: A recurrent architecture with Long Short-Term Memory units respects temporal ordering by using past observations to represent history.Each LSTM unit updates an internal state that maintains a nonlinear representation of previously observed data.
- Model architecture: The model uses LSTM units followed by rectified linear units and a softmax output to produce a probability distribution for the next price move.The LSTM design is intended to encode temporal sequences efficiently.
- Model architecture: The network predicts the next price move from the current order-book state and an internal representation of prior order-flow history.The internal state can blend nonlinear processing with history-dependence, allowing past states to affect future price dynamics.
- Training: The network parameters are learned with supervised learning by minimizing a regularized negative log-likelihood using stochastic gradient descent.Training is computationally intensive because the network contains up to hundreds of thousands of parameters and uses large datasets.
- Training: Asset-specific models are compared with a universal model trained on combined data from all stocks without specific normalization.The universal model is distributed across GPU nodes for large-scale asynchronous stochastic-gradient training.
3 Results
Deep learning reveals nonlinear, universal, stationary, and history-dependent relations between order-book dynamics and subsequent price moves. Pooling data across stocks improves forecasting and generalizes to unseen assets and later periods.
- Forecasting setup: Price-move models predict whether the next irregularly spaced mid-price change is up or down from limit-order-book state variables and their history.The interval between price moves ranges from fractions of a second to seconds.
- Path-dependence: Including longer price and order-flow histories substantially improves accuracy, providing evidence of path-dependence and possibly long-range dependence.Price dynamics depend on history, not only on the current or recent order-book state.
- Nonlinearity: Deep learning substantially outperforms linear models, increasing accuracy by 5% to 10% for most stocks.The neural network captures nonlinear relationships between price dynamics and visible order-book supply and demand.
- Nonlinearity: Around 10%, the universal nonlinear model improves forecasting accuracy relative to stock-specific linear models.This comparison uses out-of-sample results across approximately 500 stocks.
- Order-book features: The model uncovers stable nonlinear order-book features: larger best-ask depth makes the next price decrease more likely.The analysis also examines the influence of deeper order-book levels.
- Universality across assets: The universal model consistently outperforms stock-specific models, indicating forecasting-relevant common features across stocks.Features learned from one stock may help forecast price moves for another.
- Pooling and normalization: Standard normalization and sector or tick-stock partitioning do not improve training results, whereas pooling data helps especially for stocks with less data.The universal model is less exposed to overfitting because it interpolates across the richer scenario space of pooled data.
- Universality across assets: The universal model generalizes to unseen stocks: training on stocks 1–464 still outperforms stock-specific models on stocks 465–489.Its performance on the held-out stocks is roughly the same as that of a model trained on the entire dataset.
4 Discussion
The discussion presents a deep-learning model that captures universal price-formation features across stocks and time. Its generalization and use of long order-book histories support pooling broad financial data for intraday modeling.
- The model is stable across stocks and sectors, and pooled training outperforms stock-specific models, including for stocks absent from training.These results indicate that the learned features are not stock-specific.
- Model performance remains stable across time, including a year beyond the training period.
- Including order-flow history extending several hours improves prediction performance, providing evidence of long memory in price formation.
- The model extrapolates well to stocks excluded from training, including recently listed instruments or those with incomplete or short histories.
- Training one universal model is orders of magnitude less complex and costly than training or estimating thousands of single-asset models.