Stochastic Price Dynamics in Response to Order Flow Imbalance — Evidence from CSI 300 Index Futures¶
Authors: Chen Hu, Kouxiao Zhang Institution: Guolian Futures Ltd, Shanghai Year: 2025 (May 2025) arXiv: 2505.17388 Categories: q-fin.MF, q-fin.CP, q-fin.TR
Plain-language abstract¶
The Cont–Kukanov–Stoikov OFI framework was designed with symmetric windows: compute OFI over the last 5 seconds, predict the next 5 seconds. This paper argues that is the wrong way to think about it. OFI is better modelled as a step shock to the market, and the correct question is "how does price respond to this shock over varying horizons?". The authors replace the Hawkes-process framing (self-exciting arrivals) with an Ornstein–Uhlenbeck process driven by Lévy jumps — capturing both memory and heavy-tailed shocks — and embed it as the drift term in a geometric Brownian motion for price. From this they derive closed-form expressions for the log-return mean and variance, plus a quasi-Sharpe / response ratio that quantifies the drift-vs-diffusion trade-off as a function of horizon. Empirical validation uses one year of CSI 300 Index Futures tick data (≈6M ticks); OFI–price correlation rises monotonically from 0.20 at 0.5s to a stable ~0.50 plateau from 10s onward.
Key contributions¶
- OFI as a shock, not a window-matched contemporaneous signal. The paper treats accumulated OFI as a step input and studies asymmetric future response over varying horizons. This surfaces horizon dependence that window-symmetric analysis hides.
- OU-Lévy model of OFI response. Rather than modelling OFI arrivals as a Hawkes process (the dominant approach in recent literature), the authors argue the aggregated OFI signal has memory + mean-reversion characteristic of an OU process. Heavy-tailed empirical distributions motivate driving the OU with a jump-type Lévy process rather than Brownian motion.
- Modified GBM with OU drift. The canonical \(dS_t = \mu S_t dt + \sigma S_t dW_t\) becomes \(dS_t = X_t S_t dt + \sigma S_t dW_t\) where \(X_t\) is the OU-Lévy OFI response process. Coupled SDE system solved analytically.
- Response ratio (quasi-Sharpe). Explicit time-varying metric quantifying the tradeoff between OFI-driven deterministic drift and stochastic diffusion; acts as a trading-efficiency score as a function of response horizon.
- Regime taxonomy & indicator screening protocol. Monthly decomposition shows some months are "efficient" (hard to trade) while others are "inefficient" (HFT-tradable). A robust microstructure indicator should show only quantitative, not qualitative, changes across regimes — the paper proposes this as a screening criterion for new signal candidates.
- Indicator matching. Optimal pairing between OFI and other indicators (TI, Lambda, AvgEn) depends critically on forecast horizon, contradicting the common practice of examining indicators in isolation.
Method summary¶
Data¶
- CSI 300 Index Futures, tick data from the exchange (500 ms snapshot cadence).
- ~6 million ticks over 1 year.
- Resting-period handling: first/last segments of each trading session excluded; event counter \(N(t_k)\) reset at each new session.
Metrics studied¶
| Metric | Formula / idea | Source |
|---|---|---|
| OFI | \(\sum_n e_n\) with \(e_n\) signed queue-contribution (Cont–Kukanov–Stoikov 2010) | price-impact-order-book-events |
| TI (Trade Imbalance) | \(\sum_n \omega_n\) with \(\omega_n\) signed trade contribution via Lee–Ready-style classification | price-impact-order-book-events |
| Lambda | \(\lambda_k = \Delta p / v\) — price impact per unit volume (high-low range / volume) | [10] in the paper |
| AvgEn | Differential of average \(e_n\) over a window — local trend in per-event contribution | paper §2.1.5 |
OU-Lévy OFI response model¶
Aggregated OFI over a historical window is treated as a step input. The market's response \(X_t\) is modelled as an OU process driven by a Lévy process \(J_t\):
Price follows modified geometric Brownian motion with this process as drift:
The log-return \(R_t = \ln(S_t / S_0)\) has explicit mean \(\mathbb{E}[R_t]\) and variance \(\text{Var}[R_t]\) derived in Appendix A (closed-form integrals of the OU process moments).
Response ratio¶
Analogue of Sharpe under OFI-triggered trading:
This is time-varying: it rises as the deterministic OFI drift accumulates, peaks, then decays as diffusion dominates. The peak location defines an optimal holding horizon conditional on the OFI shock.
Main empirical results¶
OFI–price correlation (1-year CSI 300, Table 2.1)¶
| Horizon | 0.5s | 1s | 2s | 5s | 10s | 20s | 30s | 1m | 2m | 5m | 10m | 20m | 30m | 1h |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OFI | 0.20 | 0.28 | 0.38 | 0.46 | 0.50 | 0.51 | 0.50 | 0.49 | 0.49 | 0.34 | 0.51 | 0.52 | 0.52 | 0.54 |
| TI | 0.20 | 0.10 | 0.02 | -0.04 | -0.05 | -0.03 | 0.00 | 0.07 | 0.14 | 0.20 | 0.18 | 0.15 | 0.15 | 0.12 |
| Lambda | -0.02 | -0.05 | -0.14 | -0.41 | -0.45 | -0.45 | -0.41 | -0.25 | 0.34 | 0.35 | 0.35 | 0.33 | 0.32 | 0.29 |
| AvgEn | 0.02 | -0.01 | -0.01 | -0.01 | 0.00 | 0.00 | -0.02 | -0.05 | -0.04 | 0.00 | 0.04 | 0.07 | 0.08 | 0.11 |
- OFI is the dominant signal and remarkably stable across horizons, plateauing at 0.50–0.54.
- TI (trade imbalance) flips sign and is weak at short horizons — confirms CKS's finding that trade-only metrics are noisier than OFI.
- Lambda shows regime change around 1–2 minutes: negative at short horizons (high-impact moves anti-correlate with recent returns, i.e., mean reversion) then flips positive (momentum).
- AvgEn is effectively uninformative at any horizon — a weak metric by the paper's screening criterion.
Horizon-dependent indicator matching¶
Optimal pairing of OFI with auxiliary metrics varies with target horizon; no single "best" combination dominates. Screening criterion: robust indicators keep the same sign and statistical structure across monthly regimes.
Market regime classification¶
Monthly analysis partitions trading periods into: - Efficient regime — market prices are tight; indicators show weak predictive power; HFT unprofitable. - Inefficient regime — indicators show strong predictive structure; market is mis-pricing transiently; HFT participation helps restore efficiency.
Limitations¶
- Single instrument: CSI 300 Index Futures only. Chinese A-share cash equities, commodity futures (CFFEX, SHFE), and cross-market robustness not tested.
- 500 ms snapshot cadence is coarse relative to LOBSTER-grade tick-by-tick data — some event ordering is aggregated away.
- Closed-form results rely on OU-Lévy parameter stability within the analysis window; the paper notes regime switches but doesn't extend the SDE to regime-switching coefficients.
- No trading simulation / P&L validation — response ratio is a theoretical trading-efficiency measure, not a backtested strategy.
- Comparison with Hawkes approach is conceptual, not empirically run head-to-head on the same data — would be a natural next test (cf. forecasting-high-frequency-ofi which uses Hawkes on NSE).
Connections¶
- Chinese-markets sibling to price-impact-generalized-ofi (Su et al. 2021) — that paper ran on CSI 500 with a log-GOFI formulation; this one runs on CSI 300 Futures with an OU-Lévy modelling perspective. Together they form the Chinese-market OFI pair.
- Contrasts the Hawkes framing of forecasting-high-frequency-ofi and order-flow-filtration — same aim (model post-OFI response) but via a mean-reverting diffusion with Lévy jumps rather than a self-exciting point process. Future work could benchmark the two head-to-head.
- Inherits OFI + TI definitions from price-impact-order-book-events — this paper uses them as is but studies their responses asymmetrically.
- Regime-aware screening connects conceptually to order-flow-filtration's diagnostic-ladder methodology — both seek criteria for "what makes a microstructure indicator robust?"
- Quasi-Sharpe response ratio provides a primitive that could feed directly into execution strategies — links to optimal-execution and mpc-trade-execution.