Dhaka-Stock-Exchange-EoD-Dataset-Metadata

Temporal Coverage Bias in Financial Panel Data

Coverage-Aware Structuring Framework with Evidence from the Dhaka Stock Exchange

This repository contains the code, dataset structure, and experimental pipeline used in the study:

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

Submitted to Financial Innovation (Springer).

Overview

Financial panel datasets frequently align multiple financial instruments across a shared calendar. However, instruments are listed at different times, resulting in heterogeneous observation windows.

Naively extending time series backward to match a common calendar introduces temporal coverage bias, which can distort statistical estimates of volatility and risk.

This project proposes a coverage-aware dataset structuring framework that explicitly encodes instrument availability through an availability matrix and evaluates the statistical consequences of naive temporal alignment.

Key Contributions

• Introduces a coverage-aware dataset structuring framework for financial panel datasets.
• Formalizes temporal coverage bias arising from naive temporal alignment of financial instruments.
• Demonstrates that naive temporal padding can distort volatility estimates by:
  • ~20% reduction in return volatility
  • ~26% distortion in conditional variance estimates.
• Provides reproducible experiments using ARIMA and GARCH models.

Dataset

The dataset contains end-of-day trading records for instruments listed on the Dhaka Stock Exchange (DSE).

Coverage:

• Period: October 2012 – January 2026
• Instruments: 486
• Asset classes:
  • equities
  • treasury bills
  • mutual funds
  • bonds

Two dataset versions are provided:

1. Unadjusted dataset – raw historical price records
2. Adjusted dataset – incorporates corporate action adjustments

Coverage-Aware Representation

Instrument availability across time is encoded using an availability matrix:

A(i,t) ∈ {0,1,2,3}

Where:

0 = no observation available
1 = observation available in adjusted dataset only
2 = observation available in unadjusted dataset only
3 = observation available in both datasets

This representation preserves heterogeneous listing windows and avoids artificial temporal padding.

Experimental Pipeline

The experimental evaluation compares two dataset constructions:

1. Coverage-aware dataset
2. Naively aligned dataset with temporal padding

Steps:

1. Construct dataset representations
2. Compute log returns
3. Fit ARIMA models for illustrative analysis
4. Estimate conditional variance using GARCH models
5. Compute distortion metrics between the two constructions

Key Findings

Across 53 instruments:

• Mean return volatility distortion ≈ 20%
• Mean conditional variance distortion ≈ 26%

These results demonstrate that naive temporal alignment can significantly bias volatility estimates.

Reproducibility

Install dependencies:

pip install -r requirements.txt

Run experiments:

python scripts/arima_single_demo.py
python scripts/experiments/coverage_vs_naive.py
python scripts/experiments/cross_instrument_arima.py
python scripts/experiments/multi_instrument_volatility_robustness.py

Generate figures:

python scripts/generate_figures.py
python scripts/generate_distortion_plots.py

And any other .py files associated in the repository should run easily. The codes should also run seamlessly in anaconda.

License

This repository is released for academic and research use.