Integrated Cost-Reallocation Framework for Low-Cost AI Implementation in Regulated Finance
Description
This replication package supports the study “Integrated Cost-Reallocation Framework for Low-Cost AI Implementation in Regulated Finance.” It documents the empirical evidence used to examine how declining model-access costs affect the feasibility of AI adoption, how regulatory and organizational requirements redistribute implementation costs, and how investors respond to verified DeepSeek implementation disclosures by Chinese listed financial institutions. The package contains provider-economics observations drawn from official DeepSeek technical and pricing materials; coded evidence from six Chinese policy and regulatory instruments; and an institutional register covering 29 Shanghai- and Shenzhen-listed financial institutions, plus one sensitivity case. Event-level records include original Chinese-language quotations, English translations, source-quality assessments, timestamps, deployment architecture, implementation depth, confound screening, and coding explanations. The market-analysis files include matched-control assignments, daily A-share returns, market and China A-share FF3-style factor series, firm-level abnormal returns and cumulative abnormal returns, unique-event-date portfolios, and inferential outputs. The analyses cover the [0,0], [0,+1], and [−1,+1] event windows and incorporate cluster-preserving permutation tests, source-quality restrictions, exclusion of the largest shared-date cluster, leave-one-date-out tests, alternative timing conventions, sensitivity-case inclusion, and ±10-trading-day placebos. Provider cost and pricing observations are reproduced as reported or independently verified from official source values. The reported USD 5.576 million figure represents specified DeepSeek-V3 training stages under the technical report’s pricing assumption and should not be interpreted as total development expenditure or institutional implementation cost. The package preserves source-level evidence, coding decisions, analytical inputs, and reproducibility checks needed to evaluate and reproduce the study’s principal findings.
Files
Steps to reproduce
1. Download the replication package and open the workbook in Microsoft Excel or compatible spreadsheet software. Begin with **01_Read_Me**, which describes the file structure, variable definitions, coding conventions, event windows, and analytical specifications. 2. Use **02_Provider_Econ** to verify the DeepSeek training-cost and API-price observations against the cited official sources. Independently calculated fields reproduce the arithmetic derived from reported parameters and pricing assumptions. 3. Use **03_Policy_Evidence** to review the six Chinese policy instruments and reproduce the mapping of their provisions into data and privacy management, cybersecurity and infrastructure control, algorithm validation and accountability, and organizational preparation. 4. Use **04_Events** to reconstruct the institutional sample. Apply the stated inclusion criteria, retain each institution’s earliest verifiable implementation disclosure, assign the affected trading day using the recorded timestamp convention, and reproduce the deployment-architecture, implementation-depth, source-quality, and confound classifications from the preserved quotations and translations. 5. Use **05_Controls** to reproduce the matched-control analysis. For each primary event, select three non-announcing institutions from the same financial subsector, prioritizing exact ownership-category matches and then the smallest distance in logarithmic circulating market capitalization. 6. Use **06_Daily_Returns** and **07_Factors** to reproduce expected and abnormal returns. Estimate the market and China A-share FF3-style models over days −130 through −11 and calculate abnormal returns and CARs for [0,0], [0,+1], and [−1,+1]. 7. Compare the resulting firm-level estimates with **08_Event_CARs**. Average firm-level CARs within each affected trading day and assign equal weight to the resulting date portfolios, as documented in **09_Date_Portfolios**. 8. Use **10_Results** to verify the reported descriptive statistics, confidence intervals, conventional tests, and 10,000-draw cluster-preserving permutation results. Reproduce the robustness checks by excluding the largest event-date cluster, restricting the sample to high-quality sources, adding the sensitivity-only case, removing each event date in turn, applying the alternative Qingdao Rural Commercial Bank timing convention, and shifting events by ±10 trading days.
Institutions
- Ca' Foscari University of VeniceVeneto, Venice