Temporal Coverage Bias in Financial Panel Data: A Coverage-Aware Structuring Framework with Evidence from the Dhaka Stock Exchange

Battle Count: 55

Highly relevant to quantitative trading as temporal coverage bias directly affects volatility estimation, which underpins risk management (VaR, CVaR), portfolio construction (mean-variance optimization, risk parity), derivative pricing, and factor model estimation. The paper demonstrates that naive panel construction suppresses volatility by ~20% (return STD) and ~26% (GARCH unconditional variance) on average, with effects present in >90% of instruments. This systematically understates risk for recently-listed securities, leading to mispriced derivatives, understated VaR, and over-weighted portfolio allocations. The coverage-aware framework provides a practical, reusable correction applicable to any trading strategy relying on panel-formatted data. However, the empirical evidence is from an emerging market (DSE) and the paper focuses on data construction rather than trading strategy development directly.

The coverage-aware framework is conceptually simple and low-complexity to implement. The core correction requires: (1) recording each instrument's first and last trading dates as metadata, (2) constructing a binary availability matrix A_{i,t}, and (3) restricting analysis to dates where A_{i,t} > 0. This can be implemented with basic pandas operations (reindex with date filtering, boolean masking). The availability matrix is a standard sparse or dense boolean array. No complex algorithms or specialized infrastructure are needed. The main effort lies in data curation and ensuring listing dates are accurately recorded, not in the computational implementation itself.

Dataset is publicly available on Mendeley Data (DOI: 10.17632/23553sm4tn.4). Analysis code is available on GitHub. The availability matrix formulation is clearly defined with mathematical notation. All parameters, model specifications (ARIMA(1,0,1), GARCH(1,1)), and selection criteria (first observed after 2016, at least 400 trading days) are explicitly stated. The framework is described as reusable and analysis-agnostic.