Image

Quants Corner | FloAlpha: A Multi-Factor Model for US and European Equities

Where fund managers put their money, and how that changes, says a lot about where stocks are heading next. EPFR’s Stock Flows and Allocations dataset, launched in 2016 and re-released in September 2023, covers more than 23,000 securities. The current holding value of these securities reached above $25 trillion as of September 2026, a figure that has grown substantially since inception as new stocks are added and existing holdings grow.

The flows and holdings of this database are instrumental in deriving proof-of-concept factors. EPFR’s Quantitative Research Team has developed a suite of 14 unique factor variables, which can be expanded to 70 factors when looked across various filters. The following research goes into detail on how we can combine these factors – which all aim to capture different thematics – into one model.

FloAlpha is the name EPFR uses for its family of multi-factor models, each built from factors developed in-house by our Quantitative Research team. We have developed FloAlpha models across several asset classes and indices. The models in this Quants Corner aim to find alpha in two widely followed developed-market equity universes, the Russell 1000 and the STOXX Europe 600, and were developed together using a single method.

Each combine seven fund-flow and fund-positioning factors, selected on the stability of their returns: a core of five factors shared by both markets, plus two tilts chosen for each market. Out of sample, from September 2023 to December 2025, the sector-neutral top-minus-bottom quintile portfolio earned an annualized return of 8.0% in the Russell 1000 and 7.1% in the STOXX Europe 600, with Sharpe ratios of 1.40 and 1.48. Over the full history, from October 2011, the figures are 4.9% (Sharpe 0.85) and 4.5% (0.91).

Finding the right factor(s)

EPFR Stock Flows turns our fund-level data into stock-level signals by combining two inputs: daily flows for global mutual funds and ETFs, and those funds’ monthly security-level holdings. Allocating each fund’s daily flow across its latest holdings estimates the money moving into and out of each stock every day, while the holdings show how fund positioning shifts from month to month.

From these inputs, EPFR builds 14 factor variables:

  • Seven daily factors capture where money is moving and who is receiving it:
    • Flow Momentum (FloMo): net dollar flow into a stock, normalized by the total assets held in that security, so it shows where money is going.
    • Smart Money (FloTrend, FloDiff, FloDiff2, ActWtTrend, ActWtDiff, ActWtDiff2): track whether new money is reaching funds that are increasing or holding bullish positions in a stock.
  • Seven monthly factors capture how fund managers are positioned:
    • Manager Momentum (AllocMo, AllocTrend, AllocDiff): measures whether managers are actively adding to or cutting their allocations to a stock, with both start and end AuM used to strip out price moves.
    • Ownership Dispersion (Herfindahl, FundCt): how broadly a stock is owned. Herfindahl is based on the Herfindahl-Hirschman Index, a standard measure of concentration, and FundCt counts the funds holding the stock. Research suggests concentrated ownership is associated with higher risk.
    • Conviction (AllocSkew): stocks where a few holders take outsized positions relative to peers, signaling conviction.
    • Share Surprise (ShsSurpRev): the part of funds’ holding changes that their flows do not explain, which likely reflects other portfolio adjustments such as index rebalancing; used in reversed form.

Each variable can be computed on a subset of funds. For this analysis, we used five fund-type filters: active funds, mutual funds, passive funds, ETFs and sector funds. Fourteen variables across five filters give 70 candidate factors. Daily factors are published at T+1 5pm ET and monthly factors with a 26-day lag.

A defining feature of the model is that it uses only fund-type factors and deliberately excludes the aggregate (all-fund) factors. Each fund group, whether active, passive or other specialist, carries its own distinct signal, and pooling all funds into a single aggregate measure blurs those differences and dilutes the information that makes each group valuable. Keeping the groups separate preserves these signals and lets each one contribute to the model on its own terms.

What’s in the model

Many of the 70 candidates share their construction and are highly correlated. Variables built from the same inputs move together (the three allocation-momentum measures behave almost as one factor, as do Herfindahl and FundCt), and the passive and ETF versions of a factor are often close twins. Treating them as independent would concentrate the model in a few underlying signals. We therefore grouped the candidates into 14 themes based on their construction and correlation in both markets. The model takes at most one factor from each theme.

The heatmap below orders the 70 factors by family and theme, with the Russell 1000 below the diagonal and the STOXX Europe 600 above it. Each outlined square is a theme. Across both markets, the average correlation between factors in the same theme is 0.59, against 0.02 between factors in different themes. Factors in bold were selected: blue for the core, amber for the market tilts. Nine of the fourteen themes are represented in our final model.

70 fund-type factors in 14 themes: Correlations are high within themes and low across them

How the model was built

The model was trained on 143 months of data, from October 2011 to August 2023. The Stock Flows dataset was re-released in September 2023, so the out-of-sample period coincides with the live history.

Here are the steps taken to build this model:

  1. Factor returns. Each month, every stock is scored on the factor relative to its sector peers. We then measure the return of a portfolio that buys the stocks scoring above their sector average and shorts those scoring below it, with each position sized by how far the stock’s score sits from the average. These monthly factor returns are the basis for every step that follows.

Scores are sector z-scores, with missing values set to zero; returns are measured relative to the equal-weighted universe.

  1. Measure consistency. For every rolling three-year window in the training period, we check whether the factor made money on average. Its stability score is the share of windows in which it did: a factor that made money in 81 of the 108 windows, for example, scores 0.75. This rewards steady performance over a few large gains.

108 rolling 36-month windows, October 2011 to August 2023 , calculated separately for each market.

  1. Pick one factor per theme. For each theme, we kept the factor with the highest stability averaged across both markets. Near-ties, within 0.01, went to the factor with the higher training Sharpe ratio.
  2. Build the shared core. The 14 factors from step 3 are ranked on their average stability across both markets, and the top five that also made money in both markets during training form the core, common to both models.
  3. Tilts. Each market then adds up to two factors from the nine themes outside the core. For these themes, each market repeats step 3 using its own stability scores and Sharpe tie-breaks and adds the two most stable factors with a positive training Sharpe ratio in that market. Each model therefore has seven factors: five shared and two tailored to its market.
  4. Weights. Steadier factors get more weight. Each factor’s weight is set in inverse proportion to its volatility, then rounded to 10%, 20% or 30%. For instance a factor half as volatile as another therefore gets twice the weight.

Volatility, σ_k, is the standard deviation of the factor’s monthly returns from step 1 over the training period, calculated separately for each market. Raw weights are w_k = (1/σ_k) / Σ_j (1/σ_j), i.e. each factor’s inverse volatility as a share of the total across all seven factors, before rounding.

To construct the FloAlpha multi-factor model, each factor is compounded over its lookback period and z-scored within sectors. A stock’s FloAlpha score is the weighted sum of its seven factor z-scores. Five of the seven factors are shared by both markets, with daily flow factors carrying 60% of the weight in the Russell 1000 and 70% in the STOXX Europe 600, and monthly positioning factors making up the rest.

Putting it to the test

Each month, stocks are ranked on the FloAlpha score within their sector. The backtest holds an equal-weighted long position in the top quintile and a short position in the bottom quintile, rebalanced monthly. There is no look-ahead: factor values and index membership are taken at the end of month t−1, with each factor’s publication lag already applied, and are tested against returns in month t. The model was built on data up to August 2023 and then frozen, therefore September 2023 to December 2025 serves as an out-of-sample test. The tables below show results for the training period, the out-of-sample period and the full history.

Annual return of the top fifth of stocks minus the bottom fifth, before trading costs (October 2011 to December 2025). All figures are hypothetical backtest results before transaction costs. Past performance does not guarantee future results.

Both ends of the ranking add value in every period and in both markets: the top quintile beats the index average, and the bottom quintile lags it throughout, with most of the spread coming from these two extremes. In STOXX Europe 600, the two ends contribute roughly equally, and returns fall steadily from Q1 to Q5. In Russell 1000, nearly 60% of the spread comes from the bottom quintile, and much of that is the work of the sector-fund factors. Sector specialists know their industries well, and findings show the stocks they are reducing exposure to tend to underperform their sector peers. FloDiff2Sec, which captures this signal, is the most consistent of all 70 fund-type factors at ranking underperformers into Q5 across both Russell 1000 and STOXX Europe 600.

The model also held up on new data. Out of sample, from September 2023 to December 2025 (28 months), the model delivered Sharpe ratios of 1.40 in Russell 1000 and 1.48 in STOXX Europe 600, stronger than the 0.74 and 0.81 achieved over the training period. In the US, the improvement came mainly from the top quintile, whose annual return rose from 1.6% in training to 4.5%. Over the full history, the model made money in 12 of 15 calendar years in the US and 14 of 15 in Europe. The 28-month period is long enough to show the model working on unseen data, but too short to treat that stronger performance as the expected long-run level.

Can the results be trusted?

Pairwise cross-sectional correlations between the selected factors are low, with almost all below 0.1 in absolute value. Each factor therefore captures a unique sentiment, so the composite is diversified across signals. This is consistent with the theme-based selection: taking at most one factor per theme was designed to avoid concentrating the model in a few correlated signals, and the near-zero correlations suggest it worked.

Two pairs in the chart below are exceptions. In Russell 1000, FloDiff2Sec and FloDiff2Pas correlate at 0.35. They apply the same flow measure to two different investor groups (sector and passive), so the overlap is expected, and the modest level suggests each group still adds its own information. In both markets, AllocDiffPas, a monthly allocation-momentum measure, correlates with the price momentum at about 0.6, so part of its contribution likely overlaps with a standard momentum exposure. We accept this as the factor has a low contribution to the final model (10%) but proves to be additive.

The remaining factors show little relationship with each other or with price momentum, so most of the composite is not a repackaged momentum bet.

EPFR factor correlation files, Russell 1000 and STOXX Europe 600. Price momentum (PrcMo) is shown for reference.

Bottom line

FloAlpha’s key strength is that it combines daily flow factors, which track where money is moving, with monthly positioning factors, which show how managers are allocating, across distinct themes. The result shows that fund-flow and positioning data, broken down by investor type, can support a stock-selection model that holds up on unseen data. Out of sample, it delivered Sharpe ratios of 1.40 in the Russell 1000 and 1.48 in the STOXX Europe 600. Its factors are largely independent, and sector-fund signals stand out at spotting stocks to avoid. The results support our core idea: value runs through every layer, from the underlying flow and positioning data to factors, capturing distinct themes and specific investor groups rather than total flows, to a model that combines them. We will continue to track the FloAlpha model performance as its out-of-sample history grows.

Tags