By combining institutional and retail flow signals with volatility conditioning, our machine learning models achieved Sharpe ratios up to 1.66 and Sortino ratios up to 2.69 across multiple investment horizons—proving that participant-classified order flow contains durable, actionable alpha for systematic strategies.

Study highlights

• 3% Annual Returns on Short-Term Strategy – The 1-day horizon strategy delivered 13.0% annualized returns at 7.55% volatility (Sharpe 1.66, Sortino 2.69), compounding to nearly 300% over 2014-2025 with minimal market beta exposure (0.06), demonstrating pure idiosyncratic alpha generation from flow-based signals

• 4.11 ICIR Through Volatility Enhancement – Baseline institutional net flow signals showed an Information Coefficient of 1.12% (ICIR 2.26), but conditioning on elevated market volatility and scaling by idiosyncratic volatility more than doubled predictive power to IC 2.37% and ICIR 4.11, revealing that flow dislocations are most pronounced during market stress

• Robust Performance Across Multiple Horizons – Walk-forward validated strategies spanning short (1-day), medium (10-day), and long-term (60-120 day) horizons all demonstrated positive risk-adjusted returns with Sharpe ratios from 0.83 to 1.66, confirming that participant-classified flows contain persistent predictive information at multiple frequencies suitable for diverse investment mandates

About Flows

• Institutional flows contain a strong, consistent signal about where stocks are headed next.

• Retail flows—often dismissed as "dumb money"—also contain predictive power, achieving ICIRs up to 1.33 in S&P 500 universe.

• More surprisingly, retail and institutional flows show virtually zero correlation (-0.01 to -0.46), meaning they capture completely different market dynamics.

Short, Medium, and Longer Term Strategies Results

Short-term strategies (1-day holding):

• 13.0% annual returns with 7.55% volatility

• Sharpe ratio of 1.66, Sortino ratio of 2.69

• Nearly 300% cumulative returns over the test period

Medium-term strategies (10-day holding):

• 5.27% annual returns with 6.40% volatility

• More manageable turnover for institutional implementation

• Consistent performance across market regimes

Longer-term strategies (60-120 day holding):

• 6.8% annual returns with 7.0% volatility (60-day example)

• Minimal turnover (5% daily) suitable for large-scale deployment

• Out-of-sample Sortino ratios ranging from 1.22 to 2.7 across the full 1-120 day spectrum

Summary

Predicting stock returns from order flow is challenging because traditional market data aggregates all participants together. However, institutional and retail investors trade with different objectives, constraints, and information—creating distinct behavioral patterns that, when isolated, contain predictive value.

The research validates this by demonstrating that institutional and retail flows are effectively orthogonal (correlations from -0.01 to -0.46), confirming they provide independent information channels. Institutional flows exhibit strong mean-reversion during high-volatility regimes, while retail flows show predictive power during calm markets—reflecting fundamentally different behavioral drivers.

The models were rigorously tested using walk-forward validation with non-overlapping training, validation, and testing windows spanning 2007-2025.

LightGBM gradient boosting models were trained on comprehensive flow features including directional flows, intensity measures, participation rates, and cumulative footprints.

Performance remained consistent across multiple market regimes including the Global Financial Crisis, COVID-19, and the 2022 rate-hike cycle.

Most importantly, all strategies demonstrated minimal systematic risk exposure (betas near zero) with returns derived primarily from idiosyncratic, flow-driven sources.

This confirms the signals capture genuine cross-sectional information rather than disguised factor exposures, making them valuable additions to diversified systematic portfolios.

There is statistically significant evidence that participant-classified order flow data contains durable alpha across short-, medium-, and long-term investment horizons.

About XTech Flow™ US Equity Flow Analytics

Powered by Exponential Technology and based on LSEG data, XTech Flow™ US Equity Flow Analytics utilizes the US Consolidated Feed to apply deep high-frequency trading knowledge.

This identifies the direction of active risk-taking by institutional buy-side, market makers, and retail traders. With unprecedented 1-minute granularity and 15 years of history, the dataset provides a unique ability to distinguish institutional and retail flow, offering near-real-time market intelligence across the entire US equity market.