What can a successful model actually tell us?

3 papers · Suggested reading sequence

The manuscripts.

1. 01

   The Validation Bottleneck: Alpha Discovery When Hypotheses Are Free

   Discovery under multiple testing: why producing more hypotheses does not produce more independent evidence.

   Preprint · 2026 · Abstract, PDF & citation ↗

   Read the abstract

   Large language models have made investment hypotheses effectively free, moving the binding constraint on alpha discovery from generation to validation. Restricting a model’s data feed does not restrict its knowledge: it has read the anomaly literature, so its training cutoff is the only valid out-of-sample boundary. Put the two facts together and the strategies tested multiply without limit while the evidence that can judge them accrues at one year per year. Three consequences follow. The smallest alpha an allocator can certify rises with the logarithm of the number of strategies the industry tests, because the luckiest fluke among them looks better the more are tried. Alpha that decays faster than the evidence accrues cannot be certified within its own lifetime. And firms drawing hypotheses from the same model correlate their discoveries, on that model’s most celebrated ideas, before either holds a position. Two pre-registered experiments then measure the contamination and the bind. Told to propose predictors not yet in the literature as of a past date, five models across three lineages return that literature anyway: 26–55% of their “novel” proposals match predictors already in the library. Convergent rediscovery could explain that; it cannot explain the sign. The predictors they propose most often are those that performed worse afterward (combined z = −3.88). They reproduce fame, not future returns. Evaluated strictly after verified cutoffs, none of the roughly 900 evaluable strategies clears false-discovery control, and no model’s best pick beats the best of an equal number of random draws from the same restricted language; two of the three do worse. The models’ own stated predictions about their mechanisms held 39% of the time, against 70% for mechanisms already documented as true put through the same tests—and 46% against 86% on the one window long enough to test structure. By the second consequence, no window available today could certify a true strategy either: the nulls measure the window’s width, not the models’ quality. Hypotheses are free, and worth what they cost until validated.

2. 02

   The AI-Native Asset Manager: Institutional Cognition as Competitive Advantage

   An organizational argument about generating, challenging, sizing, and learning from investment beliefs.

   Preprint · 2026 · Abstract, PDF & citation ↗

   Read the abstract

   The asset management industry faces a structural inversion. As information becomes abundant, analysis cheap, and first-pass portfolio construction partially commoditized, the scarce resource shifts from intelligence to institutional epistemology-the architecture by which a firm generates, challenges, sizes, and learns from investment beliefs. This paper argues that the next generation of dominant asset managers will not be firms with better models in the narrow sense, but firms that have built a superior epistemic operating system: a machine-augmented cognitive architecture organized around the lifecycle of belief under uncertainty. We introduce a three-engine frameworknarrative intelligence for beta, statistical intelligence for cross-sectional alpha, and causal issuer intelligence for idiosyncratic opportunities-and show how AI enables their coordination into a single adaptive capital-allocation organism. We provide a concrete organizational blueprint, define new institutional performance metrics, and identify the governance constraints required to prevent epistemic pathologies in AI-augmented firms.

3. 03

   Beyond De Prado and Cotton: Hierarchical and Iterative Methods for General Mean-Variance Portfolios

   Three methods that bring expected-return forecasts into hierarchical and regularized portfolio construction.

   Preprint · 2026 · Abstract, PDF & citation ↗

   Read the abstract

   Hierarchical Risk Parity [López de Prado, 2016] and the Schur-complement generalisation of [Cotton, 2024] are among the most widely adopted regularised portfolio construction methods, yet both are signal-blind: they solve only the minimum-variance problem and cannot accommodate an arbitrary expected-return forecast. This paper introduces three methods that incorporate alpha signals into hierarchical and regularised portfolio construction.

   HRP-$\mu$ is a hierarchical allocator that accepts an arbitrary signal $\mu$ and nests standard HRP when $\gamma = 0$ and $\mu=\mathbf{1}$. It preserves the tree-based structure of HRP while extending it beyond the minimum-variance setting. HRP-$\Sigma\mu$ strengthens this construction by replacing inverse-variance representatives with recursive local mean–variance optima, thereby using richer within-cluster covariance information at the same $O(N^2)$ asymptotic cost.

   CRISP (Correlation-Regularised Iterative Shrinkage Portfolios) is an iterative solver for $P_\gamma w = \mu$ with $P_\gamma = (1-\gamma)\operatorname{diag}(\Sigma) + \gamma \Sigma$, so that $\gamma$ interpolates between a diagonal portfolio rule and full Markowitz. At convergence, CRISP is Markowitz applied to a variance-preserving shrunk covariance—diagonal variances unchanged, off-diagonal correlations shrunk—with $\gamma$ tuned for out-of-sample Sharpe rather than covariance-estimation loss.

   In Monte Carlo experiments across multiple covariance regimes and estimation ratios, HRP-$\mu$ and HRP-$\Sigma\mu$ both outperform plain HRP—not only on general expected-return forecasts, where HRP is signal-blind, but also on the minimum-variance problem $\mu = \mathbf{1}$ that is HRP’s home territory—with HRP-$\Sigma\mu$ consistently improving on HRP-$\mu$. CRISP at intermediate $\gamma$ is the dominant method in both regimes, outperforming HRP, Cotton, Ledoit–Wolf shrinkage, direct Markowitz, and the signal-aware hierarchical methods at every sample size tested.

   Keywords: portfolio optimization, hierarchical risk parity, Schur complement, signal-dependent allocation, covariance shrinkage, out-of-sample evaluation

   JEL Classification: G11, C61, C63, C55

   Code: All code, data, and reproduction scripts are available at https://github.com/bwuebben/beyond_hrp_and_cotton.