Forecasting High Frequency Order Flow Imbalance
Forecasting High Frequency Order Flow Imbalance
This paper addresses the challenge of forecasting high-frequency Order Flow Imbalance (OFI), a key indicator of buying and selling pressure in financial markets. The authors model the arrival of buy and sell market orders as a bivariate Hawkes process, which naturally captures both self-excitation (clustering) and cross-excitation between the two order types. They develop a simulation-based method to forecast the entire distribution of OFI over a short horizon (e.g., one minute), rather than just a point estimate. To compare multiple forecasting models—including various Hawkes kernels (exponential, power-law, sum-of-exponentials, non-parametric) and linear time series models (VAR)—they employ the Superior Predictive Ability (SPA) test of Hansen (2005), which allows for fair comparison across many models while controlling for data snooping. Using tick data from the National Stock Exchange (NSE) for Nifty futures, they find that the Hawkes process with a sum-of-exponentials kernel consistently outperforms all other models in forecasting the OFI distribution. The paper also introduces a deterministic method for classifying trades as buy or sell using exchange order IDs, avoiding the errors inherent in classification algorithms like Lee-Ready. The findings have practical implications for market makers and regulators, as accurate OFI forecasts can help manage adverse selection and monitor market stability.
Highlights
- 1Develops a Hawkes process-based model for Order Flow Imbalance (OFI) that captures cross-dependence between buy and sell market orders.
- 2Introduces a simulation-based method to forecast the near-term distribution of OFI, enabling risk management for high-frequency traders.
- 3Proposes a model comparison framework using the Superior Predictive Ability (SPA) test to identify the best forecasting model among many.
- 4Applies the methodology to NSE Nifty futures tick data, showing that a Hawkes process with a sum-of-exponentials kernel outperforms alternatives.
- 5Provides a deterministic trade classification method using exchange order IDs, avoiding estimation errors common in Lee-Ready or bulk classification.
Methods
- MHawkes processes with parametric (exponential, power-law, sum-of-exponentials) and non-parametric kernels to model buy/sell trade arrival intensities.
- MSimulation-based forecasting of OFI distribution by generating future trade paths from the fitted Hawkes process.
- MSuperior Predictive Ability (SPA) test of Hansen (2005) for comparing multiple forecasting models, using a loss function based on forecast distribution accuracy.
- MRolling window estimation and forecasting over 1-minute horizons on tick-by-tick data.
Results
- RThe Hawkes process with a sum-of-exponentials kernel provides the best forecast of OFI distribution among all competing models.
- RACF and PACF of realized OFI show significant autocorrelation, motivating the use of time series models like VAR and Hawkes processes.
- RThe proposed method successfully captures self- and cross-excitation in order flow, improving forecast accuracy over models that ignore cross-dependence.
- RDeterministic trade classification from NSE data eliminates classification errors, leading to more reliable OFI estimates.
Analyze Paper
Generate insights from "Forecasting High Frequency Order Flow Imbalance".