Machine Learning the Lifecycle of Unobserved Performance
Extending Agarwal, Ruenzi & Weigert (Journal of Finance, 2024) with the ML toolkit of Bali, Beckmeyer, Moerke & Weigert (Review of Financial Studies, 2023)
Hatef Tabbakhian · github.com/Leotaby
PythonRStataC++NLP / LLM
UP spread (Q5-Q1)
0.67%/mo
Agarwal et al. (2024)
ML OOS R-squared
4.8%
Neural Net
Long-short Sharpe
0.76
ML-predicted UP
IVOL x UP peak
+0.72%/mo
Double-sort corner
Core hypothesis. Unobserved Performance is not permanent. It emerges when a manager is young, peaks around years 3 to 6, and decays as AUM grows. If true, some of the UP spread in Agarwal et al. (2024) is a lifecycle effect, and ML can predict where a fund is in its cycle.
The UP Lifecycle Curve
Hypothesized UP over fund age. Shaded = confidence band.
IVOL Effect by Fund Age
Q5-Q1 IVOL alpha spread decays with fund age.
Implication. The positive IVOL effect from Bali & Weigert (2024, RoF) should be concentrated in young funds where UP is highest. This project uses LLM-extracted textual signals from fund letters to detect lifecycle transitions.
Panel A: UP-Sorted Quintile Alphas
FH7 alpha, %/month. Agarwal et al. (2024, JF) Table 3.
Q5-Q1: 0.67%/mo (t = 4.12). About 8% annual outperformance.
Panel B: IVOL-Sorted Quintile Alphas
Bali & Weigert (2024, RoF) Table 5.
Q5-Q1: 0.63%/mo (t = 3.87). Opposite of stocks.
The gap. Both UP and IVOL predict alpha independently. Nobody has combined them with ML, or asked whether UP follows a lifecycle.
OOS R-squared by Model
Expanding-window prediction of next-month UP.
SHAP Feature Importance (GBT)
Lagged UP + IVOL + NLP sentiment dominate.
Cumulative Long-Short Returns: ML vs OLS
Long Q5, short Q1 of predicted UP. Monthly rebalance.
NLP adds value. Adding LLM-extracted features from fund letters increases GBT R-squared from 3.6% to 4.3%. The multi-agent pipeline uses separate agents for investor letters, SEC filings, and social media.
Double-Sort: IVOL x UP Quintiles → Monthly Alpha (%)
25 portfolios, FH7-adjusted. Blue = negative, white = zero, red = positive.
Alpha Matrix (%/mo)
Numerical values. Row = IVOL quintile, Column = UP quintile.
The IVOL effect strengthens with UP (Q5 col spread: +0.60 vs Q1: +0.35). High IVOL only helps when the manager has real skill.
What this shows. Top-right (high IVOL + high UP) earns +0.72%/mo. Bottom-left is negative. IVOL captures skill only when accompanied by high UP. ML models separate these populations out of sample.