Abstract
Accurate outlier detection is essential when distributions are asymmetric and heavy-tailed, a common feature of financial returns. We revisit boxplot-style rules using daily log returns of AAPL (10,319 trading days, 1984–2025; skewness ≈ −1.93, excess kurtosis ≈ 56.9). We compare the classical Tukey rule (1.5×IQR), an adjusted Medcouple approach (HV; implemented here as a fixed-count, rank-based split across tails), and a simple Quantile-Adjusted Boxplot (QAB) that sets fences at symmetric tail quantiles (e.g., 1st–99th percentiles). Beyond flag counts, we evaluate whether flags at time t anticipate short-horizon risk using two forward targets: next-day absolute return |rt+1| and 5-day forward realized volatility (annualized). We report classification metrics for top-decile |rt+1| and mean “uplift” (flagged minus unflagged) in both targets, and we assess robustness via bootstrap Jaccard similarity of flagged sets. Empirically, Tukey over-flags (≈5.24%), while QAB (≈2.02%) and HV (≈2.0–2.1% by construction) are selectively sparse and deliver larger risk uplift. Tukey attains the highest recall and the best F1/AUC but at the cost of many false positives; QAB and HV achieve substantially higher precision and stronger risk separation per alert. Stability under resampling is high for all methods (Jaccard ≥ 0.92), with Tukey most stable and HV slightly more stable than QAB. Overall, QAB provides a transparent, distribution-adaptive alternative; HV offers a similarly selective rule with a small recall edge; Tukey is preferable when recall and operational stability dominate precision.

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2025 Chao Xu (Author)