Introduction

Trading portfolios contain hundreds or thousands of assets. Forecasting each asset independently ignores common patterns: equity sectors move together, commodities respond to dollar strength, bonds respond to rate expectations. Hierarchical time-series models exploit these relationships, sharing information across assets and improving forecasts. This article explores hierarchical approaches.

Motivation for Hierarchical Modeling

When forecasting independently, each asset gets limited data. A thinly-traded stock has only 250 trading days per year of data; forecasting with little data is difficult. Hierarchical models leverage information from related assets: if this stock is in the tech sector and sector has strong uptrend, the individual stock likely participates. If tech sector is weak, stock is constrained even if company-specific news is good.

Hierarchical Model Structures

Time-Series Hierarchy

Level 0: Market level forecast (S&P 500 return). Level 1: Sector forecasts (Technology, Healthcare, Finance). Level 2: Individual asset forecasts. Forecasts are reconciled: sum of sector forecasts should equal market forecast.

Advantage: asset-level forecasts are informed by sector-level patterns. Disadvantage: reconciliation introduces constraints that might mask genuine divergences.

Cross-Sectional Hierarchy

Model relationships between assets directly. Asset returns are functions of common factors (market return, factor returns) and idiosyncratic components. Multi-factor models decompose return into systematic and idiosyncratic parts.

Equation: R_i = β_i * R_market + γ_i * Factor_1 + ... + ε_i. This separates market/factor contributions from asset-specific alpha.

Temporal Hierarchies

Forecast at multiple frequencies: monthly returns inform daily trading, quarterly earnings inform monthly momentum. Use temporal reconciliation: sum of daily returns should equal monthly return (obviously true, but ensures internal consistency).

Hierarchical Vector Autoregression (HVAR)

Extend VAR models to hierarchies. Instead of modeling each asset separately, model joint dynamics of hierarchy. Asset returns depend on own history and related assets' history.

HVAR specification: R_t = β_0 + β_1 * R_{t-1} + β_2 * R_market,{t-1} + β_3 * R_sector,{t-1} + error. This captures own momentum, market momentum, and sector momentum in single model.

Factor Models as Hierarchical Frameworks

Multi-factor models naturally encode hierarchy: assets returns are linear combinations of factors. Common factors (market, size, value, momentum) explain variation across assets. Idiosyncratic components are asset-specific.

Hierarchical extension: factor models with random coefficients. Each asset has different exposure to factors, and exposures evolve over time. Bayesian framework enables shrinkage: assets with limited data have exposures shrunk toward sector average, improving stability.

Handling Missing Data and Unbalanced Hierarchies

Real portfolios have missing data: some assets haven't traded recently, some factor data is delayed. Hierarchical models handle missingness naturally through joint estimation: even if individual asset returns are missing, sector-level information constrains them.

Unbalanced hierarchies (sectors with different numbers of assets) are handled by explicitly modeling imbalance: don't assume all sectors contribute equally, weight by number of assets.

Reconciliation Methods

Bottom-up reconciliation: forecast each asset, sum to sector and market. Simple but ignores higher-level information. Top-down reconciliation: forecast market, allocate to sectors and assets. Middle-out: forecast some levels, derive others. Optimal reconciliation: estimate all levels, reconcile using variance-covariance structure.

Optimal reconciliation minimizes forecast error across hierarchy, finding balance between respecting individual forecasts and maintaining hierarchical constraints.

Dynamic Hierarchies and Regime-Switching

Hierarchies aren't static. In normal times, assets within sector are correlated. In crises, correlations spike and individual-sector distinction breaks down. Use regime-switching models: sector hierarchy in normal regimes, market-wide hierarchy in crisis regimes.

Bayesian Approaches and Shrinkage

Bayesian hierarchical models with shrinkage priors are powerful. Assets are modeled with hierarchical priors: asset alpha comes from sector alpha plus idiosyncratic component. Sectors come from market alpha plus idiosyncratic. This hierarchy enables borrowing strength: weak assets borrow information from sector, improving estimation.

Practical Implementation

Tools: statsmodels (VAR, factor models), PyMC3 or Stan (Bayesian hierarchical), scikit-learn (PCA for factor extraction). Start simple: multi-factor model. Graduate to hierarchical VAR if asset correlations matter. Use Bayesian if limited data.

Conclusion

Hierarchical time-series models improve multi-asset forecasting by exploiting relationships across assets. Factor models provide simple frameworks; hierarchical VAR provides flexibility; Bayesian approaches enable shrinkage to shared distributions. For portfolios with hundreds of assets, hierarchical approaches are essential—they reduce model complexity, improve stability, and incorporate domain knowledge about asset relationships.