By Charlie Reese — Oct 22, 2025

Hivemind Factors Explain Incremental Out-of-Sample Variance

Why 2025 Has Been Challenging For Quantitative PMs
Creating Macro and Subindustry Hivemind Factors
How We Measured Out-of-Sample Explained Variance (R²)
Results: Hivemind Factors Explain Incremental Out-of-Sample Variance
Email Us: [email protected]

💡

Hivemind is ForecastOS' solution for distilling insight from unstructured time-series data. Read more about Hivemind here and here.

For this article, Hivemind was used to create stock-specific exposures for 1) the top macro themes and 2) nascent subindustries.

1. Why 2025 Has Been Challenging For Quantitative PMs

If you ran a large book through 2025’s volatility, you probably spent more time on the phone than in the market.

“Are you down? We’re off 3%.”

Most of those calls were really a fishing expedition for: "what exposure did I miss?" and "does everyone have it, or just me?" The uncomfortable truth is that, outside of a handful of slow-moving market, industry, and style boxes, few managers could answer that question with conviction.

The challenging 2025 market environment has been quantifiable; it has been measurably harder to isolate alpha. Out-of-sample variance explained by market, industry / subindustry, and common style factors dropped by roughly a third vs previous benchmarks. Even sophisticated hedge funds like Citadel and Millennium have had lackluster years to date.

What could be causing this?

We believe the challenging 2025 investment environment was caused in part by growth in emergent macro factors and nascent subindustries. We outline and test this hypothesis below.

2. Creating Macro and Subindustry Hivemind Factors

To better explain variance

We believe out-of-sample explained variance declined in 2025 (in part) due to an increase in A) emergent macro factors and B) nascent subindustries not captured well by traditional factors.

A) Emergent Macro Factors

A quick look at the ForecastOS Hivemind UI gives us some ideas what macro factor exposures might be to blame:

Given the above, and what we learned from our most recent article Measuring Emergent Sources of Risk & Return Since 2020, we'll use the following Hivemind stock-specific macro exposures to better capture volatility:

AI / generative AI
Defense and national security spending
Green energy
Tariffs
Medicaid cuts
Inflation

B) Identifying Nascent Subindustries

For nascent subindustries, we can create a Hivemind pipeline to describe each company in a custom, standardized way.

We can then feed this Hivemind standardized output into an unsupervised learning algorithm to systematically identify groups of similar companies. Doing so at our preferred level of specificity, we identify 292 nascent subindustries across several thousand US listed companies.

If a company does it, we've identified it, and our process is resilient to new / changing businesses going forward!

3. How We Measured Out-of-Sample Explained Variance (R²)

Using our aforementioned 6 Hivemind macro factors and our 292 new Hivemind subindustries, we can measure incremental out-of-sample explained variance.

Benchmark: The Traditional Factors We Used

To perform our analysis, we used the following traditional factors: a market factor (1 for all companies; our intercept), industry (12) and subindustry (32) factors from a major data vendor (one-hot encoded), and style factors.

To calculate our QVM (quality, value, momentum) and size style factors, we used the below ForecastOS feature engineering config:

feature_normalization:
  winsorize:
    - 0.05
    - 0.95
    - - datetime
  standardize:
    - datetime

feature_adjustments_post_normalization:
  shift: 
    - 20 # Avoid lookahead
    - - id
      - datetime
    - id
  zero_fill: []

features:
  # Quality
  
  # Quality: margin
  gross_income_margin_ltm:
    uuid: 49c61109-c029-4c4d-85c9-b1f33a72ea50
  ebit_margin:
    uuid: 47d7d1b0-46cc-4620-b0d3-35aa33fe490a
  net_income_margin_ltm:
    uuid: 7f1e058f-46b6-406f-81a1-d8a5b81371a2
  
  # Quality: margin growth
  gross_income_margin_growth_ltm_growth_over_ltm_lag_2a:
    uuid: e773a123-6799-408d-a54f-5474a7236d9d
  ebit_margin_growth_ltm_growth_over_ltm_lag_2a:
    uuid: 9a0cc931-567c-4533-ba15-7a510d03932d
  net_income_margin_growth_ltm_growth_over_ltm_lag_2a:
    uuid: 5d6bdfe1-412b-4c05-bde9-3279b53527df
  
  # Quality: leverage
  debt_total_prev_1q_to_ebit_ltm:
    uuid: 53a422bf-1dab-4d1e-a9a7-2478a226435b
    adjustments: 
      neg_to_max:
        - datetime 
    adjustments_post_normalization:
      sign_flip: []
  debt_total_to_ev_prev_1q:
    uuid: 34f37f4b-2fd9-4b23-8a72-9d421eacf45f
    adjustments_post_normalization:
      sign_flip: []
  
  # Value
  
  # Value: trading_forward
  ev_to_sales_cons_est_fwd_ntm:
    uuid: 6d1eea9d-e2a3-4332-9b81-942031ee4058
    adjustments_post_normalization:
      sign_flip: []
  market_cap_to_net_income_open_dil_over_fwd_ntm:
    uuid: ca90214e-32d7-40eb-badf-4753957c1035
    adjustments_post_normalization:
      sign_flip: []
  market_cap_to_operating_cf_open_dil_over_ocf_fwd_ntm:
    uuid: 23919e79-84ce-40e2-85a4-1c30cf4d3c51
    adjustments_post_normalization:
      sign_flip: []
  
  # Value: trading_backward
  ev_to_sales_ltm:
    uuid: e6b8dbea-62ca-4bdd-9602-f603330a65c6
    adjustments_post_normalization:
      sign_flip: []
  ev_to_ebit_ltm:
    uuid: f65f7b6a-5433-4543-8f15-5f50e43dd9f9
    adjustments_post_normalization:
      sign_flip: []
  market_cap_to_net_income_open_dil_over_net_income_ltm:
    uuid: 7e4264a9-a137-4f5b-bc10-1f1e90d2f3f5
    adjustments_post_normalization:
      sign_flip: []
  market_cap_to_operating_cf_open_dil_over_ocf_ltm:
    uuid: 5f050fce-5099-4ce9-a737-5d8f536c5826
    adjustments_post_normalization:
      sign_flip: []

  # Momentum
  
  # Momentum: return
  prev_252d_open_return:
    uuid: ed02d053-a0e1-4447-aff2-ad399f770f14
  prev_20d_open_return:
    uuid: e27ee2aa-fed9-4f85-87ef-77354791f480

  # Size
  style_size: 
    uuid: "dfa7e6a3-671d-41b2-89e3-10b7bdcf7af9" # market_cap_open
    adjustments: 
      log: []
  

features_derived:
  
  # Quality
  margin_composite:
    calculate_with: normalized
    formula: 
      mean:
        - gross_income_margin_ltm
        - ebit_margin
        - net_income_margin_ltm
  margin_growth_composite:
    calculate_with: normalized
    formula: 
      mean:
        - gross_income_margin_growth_ltm_growth_over_ltm_lag_2a
        - ebit_margin_growth_ltm_growth_over_ltm_lag_2a
        - net_income_margin_growth_ltm_growth_over_ltm_lag_2a
  leverage_composite:
    calculate_with: normalized
    formula: 
      mean:
        - debt_total_prev_1q_to_ebit_ltm
        - debt_total_to_ev_prev_1q
  
  style_quality:
    calculate_with: normalized
    formula:
      mean:
        - margin_composite
        - margin_growth_composite
        - leverage_composite
  
  # Value
  trading_comps_fwd_composite:
    calculate_with: normalized
    formula: 
      mean:
        - ev_to_sales_cons_est_fwd_ntm
        - market_cap_to_net_income_open_dil_over_fwd_ntm
        - market_cap_to_operating_cf_open_dil_over_ocf_fwd_ntm
  trading_comps_prev_composite:
    calculate_with: normalized
    formula: 
      mean:
        - ev_to_sales_ltm
        - ev_to_ebit_ltm
        - market_cap_to_net_income_open_dil_over_net_income_ltm
        - market_cap_to_operating_cf_open_dil_over_ocf_ltm
  
  style_value:
    calculate_with: normalized
    formula:
      mean:
        - trading_comps_fwd_composite
        - trading_comps_prev_composite
  
  # Momentum
  style_momentum:
    calculate_with: raw
    formula:
      subtract:
        - prev_252d_open_return
        - prev_20d_open_return

Note that we only use style_quality, style_value, style_momentum, and style_size in our regression; other factors listed above are simply inputs to the above composite QVM style factors.

Our Regression

Our regression was cross-sectional for the top 1000, 2000, and 3000 US stocks by market cap, fitting a separate model for each 20-trading-day return period. To control instability from correlated or thinly-supported factors, we applied ridge regularization. The shrinkage penalty tempered extreme coefficients, ensuring that each exposure competed for explanatory power without overfitting to noise.

Further, to prevent wild coefficients for the 292 emerging sub-industry factors, we capped their effective influence through both ridge shrinkage and scaled exposure weighting. In practice, this meant thin sub-industries were allowed to express signal - but only in proportion to the data depth supporting them, keeping the regression stable.

For each date, we split the universe into two disjoint sets: one for training (75%), where coefficients were estimated, and another for testing (25%), where regression performance (i.e. explained variance or R²) was measured. This intra-period separation helped assess generalization i.e. whether the fitted relationships held beyond the specific sample used for estimation.

4. Results: Hivemind Factors Explain Incremental Out-of-Sample Variance

💡

The incremental R² of Hivemind factors is comparable in magnitude to well researched QVM+size style factors.

As per above, Hivemind’s macro and subindustry factors incrementally improved out-of-sample explained variance meaningfully; they identified systematic risk that conventional factors missed.

Perhaps unsurprisingly given dynamic macro and market narratives as of late, the improvement was stronger in 2025 than in prior years - roughly double.

5. Email us: [email protected]

If you’re finding volatility harder to manage and alpha harder to locate, we'd love to compare notes!

In today’s fast-moving investment landscape, where generative AI, inflation fluctuations, geopolitical upheavals, and global trade shifts increasingly dictate market behaviour, static risk models fall short. Only tools rooted in perceived causality, not outdated correlations, can help institutional investors quickly and accurately surface and manage dynamic exposures.

That’s exactly what ForecastOS Hivemind delivers. By converting unstructured data - podcasts, filings, and other live feeds - into clean, point‑in‑time factor signals, Hivemind equips you with real-time insights for emergent themes.

ForecastOS Hivemind isn’t just a tool - it’s your edge. In an era where the pace and complexity of market forces are escalating, Hivemind empowers you not just to respond, but to anticipate and act with conviction.

Let’s partner to help you discover, measure, and manage emergent sources of alpha and risk!

To learn more about Hivemind, or to schedule a demo, email us at hi@ forecastos.com

Table of Contents