David Ardia

Can complex models really outperform simple ones in finance—even with small samples? This new paper explores why the answer is yes, showing that combining diverse, intricate models leads to stronger, more stable investment strategies. A must-read for anyone exploring machine learning for return prediction: https://bit.ly/45dsduZ

117

1 Comment

🚨 LLMs are not reliable tools for macroeconomic forecasting A new paper, The Memorization Problem: Can We Trust LLMs’ Economic Forecasts? (Lopez-Lira, Tang, Zhu, 2025), provides the first systematic evidence that large language models recall vast amounts of historical economic and financial data rather than truly forecasting it. Key findings: 1) LLMs can perfectly recall pre-cutoff values (GDP, unemployment, stock indices, etc.) with suspicious accuracy. 2) Even when explicitly instructed to ignore future information, models still produce forecasts anchored to memorized outcomes. 3) Masking company names or dates doesn’t help—LLMs can often reconstruct the original entities from subtle context. 4) Apparent predictive success on pre-cutoff data is often an artifact of training data exposure, not genuine economic insight. The implication is clear: any LLM-based macro or market forecasting within their training period cannot be trusted. The only rigorous way forward is testing exclusively on post-cutoff data. 📄 Full paper here: https://lnkd.in/eSBXJvrR

113

16 Comments

Schur Complementary Allocation is now part of skfolio v0.13.0. Thanks to Peter Cotton for his contribution and review! This allocation method uses Schur-complement-inspired augmentation of sub-covariance matrices to smoothly interpolate between Hierarchical Risk Parity (HRP) and the Minimum-Variance Portfolio (MVP). HRP, which leverages hierarchical clustering, tends to yield more diversified and stable allocations than MVP, which is more sensitive to estimation error and can produce concentrated weights. Schur Complementary Allocation allows us to explore the continuum between them by tuning a regularization factor that controls how much off-diagonal information is incorporated into the augmented covariance blocks. We can then leverage the scikit-learn toolkit using randomized search cross-validation to select the regularization factor that maximizes a chosen out-of-sample portfolio metric. A key implementation focus was numerical robustness, maintaining a well-conditioned covariance matrix across the regularization path. Example in the comments below. #skfolio #opensource #machinelearning #portfoliooptimization #quantitativefinance #quant #portfoliomanagement

112

7 Comments

In the new issue of Mathematical Finance: Volatility Models in Practice: Rough, Path-Dependent, or Markovian? by Eduardo Abi Jaber and Shaun Li https://lnkd.in/eAkXkd-F We present an empirical study examining several claims related to option prices in rough volatility literature using SPX options data. Our results show that rough volatility models with the parameter  are inconsistent with the global shape of SPX smiles. In particular, the at-the-money SPX skew is incompatible with the power-law shape generated by these models, which increases too fast for short maturities and decays too slowly for longer maturities. For maturities between 1 week and 3 months, rough volatility models underperform one-factor Markovian models with the same number of parameters. When extended to longer maturities, rough volatility models do not consistently outperform one-factor Markovian models. #quantitativefinance #quant #volatility

223

4 Comments