PINN-Dynamic ST-EdgeGAT and graph Péclet numbers on time-varying planar financial networks: code and data

# PINN-Dynamic ST-EdgeGAT and graph Péclet numbers on time-varying planar financial networks Code, processed data and results accompanying > G. Lim, *Advection–diffusion transport and graph Péclet scaling of volatility on time-varying planar> financial networks*, submitted to Physica A (2026). Version 2 (revision of 2026-11-03). Deposit assembled on 2026-11-03.Zenodo DOI: https://doi.org/10.5281/zenodo.23891620 ## Contents | Folder | Content ||---|---|| `code/` | Python pipeline `01`–`19` (data collection, RMT filtering, residual MST/PMFG, time-varying spectra, dynamic PMFG snapshots, PINN-Dynamic ST-EdgeGAT training, ablation / surrogate / lambda sweep, cross-market comparison, figure and table generation) and MATLAB visualisation scripts `r06`, `r08`, `r11` || `data/00_raw_metadata/` | KRX stock master list (codes, names, market) || `data/01_processed/` | Cleaned adjusted prices and log-return panel (2015-01-16 to 2026-10-30, N = 1,087), standardised returns, raw and RMT-filtered correlation matrices || `data/02_network/` | Full-sample residual correlation, Mantegna distance, residual MST (edges, GraphML), Ward clusters and node centralities || `data/03_time_varying/` | Sliding-window spectra and MST metrics (N = 400, W = 240 d, step 5 d) || `data/04_static_pmfg/` | Static residual PMFG of the 250-stock sub-universe (adjacency, edge attributes, PyG tensors) || `results/05_dynamic_pmfg_and_pinn/` | 278 causal dynamic PMFG snapshots (`KRX_Dynamic_PMFG_Snapshots.npz`), trained reference model (`best_pinn_dynamic_stgat.pt`), per-asset physics coefficients and Péclet numbers, cluster summary, test-set fluxes (`.mat`) || `results/06_ablation_null_lambda/` | Per-run JSON results and checkpoints for M1–M4 x seeds 42–46, lambda sweep, surrogate distributions, Tables 1–3 and B.1 || `results/07_cross_market/` | 150-stock return panels of KOSPI, KOSDAQ, S&P 500, NASDAQ, Nikkei 225 (2019–2026), window-wise metrics, PMFG-masked empirical Péclet ratios, Table 5 || `manuscript/` | LaTeX sources (`main.tex`, `references.tex`, `highlights.tex`, auto-generated `numbers_auto.tex` and `table_*.tex`), figures (PDF + PNG, 300 dpi) and the compiled PDF | ## Reproducing the results Environment: Python 3.11, PyTorch 2.4 (CUDA), NumPy, pandas, SciPy, NetworkX 3.6, matplotlib 3.11, pyarrow,FinanceDataReader (only for re-downloading raw prices). A `conda env create -f code/environment.yml` file isprovided as a starting point. 1. Raw daily prices are not redistributed. They can be re-downloaded with `code/01_collect_krx_daily.py` (Korea Exchange via FinanceDataReader) and, for the cross-market panels, by `code/13_...py` (Yahoo Finance). All downstream steps run from the processed files in `data/`.2. `02`–`05` rebuild the correlation, network and time-varying products in `data/`.3. `09` rebuilds the dynamic PMFG snapshots; `10` trains the reference model and writes the Péclet table (`seed 42`, deterministic on the hardware used); `12 --stage all` runs the five-seed ablation, the lambda sweep and the 500-surrogate null tests (about 2 h on an RTX 3080 Ti); `13` runs the cross-market comparison (about 30 min on 20 cores).4. `15` draws Figs. 1, 2, 3, 5; `16` exports every number used in the manuscript to `numbers_auto.tex` and the tables to `table_*.tex`; `17` checks the citation order; `18` checks Physica A format constraints.5. Compile with `pdflatex main.tex` (twice) in `manuscript/` (elsarticle class). Paths at the top of each script point to the original working directory(`C:\Users\gclim\OneDrive\문서\MATLAB\Financial_Data`); set `BASE_DIR`/`BASE` to the unpacked location. ## Key quantities (reference model, test period 2025-02-10 to 2026-10-23) See `manuscript/numbers_check.csv` for every number quoted in the paper with its macro name. ## Licence Code: MIT License. Processed data, results and figures: CC BY 4.0. Raw market data remain subject to the termsof their providers (Korea Exchange, Yahoo Finance) and are not included. ## How to cite G. Lim (2026). PINN-Dynamic ST-EdgeGAT and graph Péclet numbers on time-varying planar financial networks:code and data (v2.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23891620 ## Contact Gyuchang Lim, Kyungpook Institute of Oceanography, Kyungpook National University ([e-mail]).