Quantile regression as a sanity check on asset valuation (yegortkachenko.com)

Yegor Tkachenko proposes using quantile regression of price on time as a lightweight sanity check for asset valuation: instead of modeling mean price paths, fit conditional quantiles (e.g., 10th, 25th, 75th, 90th) so you get a family of trend lines that indicate historically “cheap” and “expensive” levels over time. This is significant for the AI/ML and quant finance communities because it produces interpretable, distribution-aware bounds that are robust to outliers and skewed returns, making it a practical complement to forecasting models or automated trading signals for flagging over- or under‑priced assets. Technically, the recipe is simple and convex: set X = [1, t] (intercept + time index) and estimate β by minimizing the pinball (tilted absolute) loss for chosen τ plus an ℓ1 regularizer (λ·||β||1) to stabilize coefficients. The pinball loss weights positive and negative residuals by τ and (1−τ) so the optimizer converges to the conditional τ-quantile; CVXPY solves this efficiently. The notebook fetches price series (e.g., from stooq), fits quantiles (0.1, 0.25, 0.75, 0.9), plots them against historical prices, and derives a simple annualized return proxy from averaged quantile slopes. Low complexity, interpretability, and robustness make this a useful operational check for ML-driven valuation pipelines.