Random Accountant Do Data Analysis

About Me

Finance professional,
data curious.

I started my career at Deloitte Vietnam as an auditor, where I spent 3.5 years working across industries. I joined as a third-year university student — running between the printing room and binding desks to get financial statements out the door, learning to read them page by page under the guidance of partners' secretaries. From there I moved into fieldwork, working directly with clients and eventually leading audit teams. Numbers came naturally to me, and more importantly, I enjoyed thinking critically about the connections between them. The last 1.5 years at Deloitte took me into IFRS and US GAAP engagements, which led naturally to a Finance Executive role at Home Credit Vietnam.

At Home Credit, I worked with a 30-million-row Oracle data warehouse — writing and extending production SQL covering IFRS 9 stage classification, provision variance analysis, and loan portfolio segmentation across multiple product types. That experience made me realise how much I enjoy working at the intersection of finance domain knowledge and data.

Today at Vinamilk, I work in cost & pricing accounting while independently building data workflows — regression-based sales forecasting, channel-to-P&L analysis, and Power BI reporting. I'm actively developing a portfolio that connects real financial datasets to data science techniques.

Outside the technical work, I think of myself as a natural problem solver — whether that's helping a colleague fix a formula in Excel, or recommending the right angle to take on an audit engagement. I genuinely believe people have more potential than they realise, and I've always found satisfaction in helping draw that out. Teaching has been a constant thread throughout my career — from university, through Deloitte and Home Credit, to today. I see it as the most effective way to both share and refresh knowledge. And honestly, I just enjoy talking to people — the interesting ones especially.

What I bring

Finance domain depth most data scientists don't have — I understand what the numbers mean, not just how to query them.

What I'm building toward

Data Scientist or Data Management roles in finance or consulting — where deep domain expertise and analytical capability work together.

Currently

Pricing & Costing Specialist at Vinamilk · Ho Chi Minh City, Vietnam · Open to opportunities.

Portfolio

Projects

Real financial data, real analysis. Projects are being built progressively — check back as new ones are added.

Live

Loan Portfolio Provision Variance & ECL Model

End-to-end IFRS 9 credit risk analytics on 3 consumer loan products (CD, TW, CLW) across 27 months (May 2020–Jul 2022), including the COVID stress period. Built a 6-state Markov chain transition model (Current → DPD1-30 → DPD31-90 → DPD91-180 → WD New → WD Aged) using balance-weighted roll-rate matrices averaged across 26 consecutive month pairs. T¹² raised to the 12th power for 12-month forward projection and cumulative PD estimation. ECL computed as PD × LGD × EAD on full outstanding (Principal + Interest + Penalty) — adding I+P increases total ECL by +137B VND (+6.8%) vs principal-only. Interest and penalty ratios validated as early warning signals: CLW penalty ratio is 33.9% higher in migration months vs stable months, and interest ratio +36.1% — the strongest combined signal across all products.

Python Markov Chain IFRS 9 ECL NumPy pandas Credit Risk Oracle SQL

Live Page → GitHub

Live

FMCG Sales Forecasting & Regression Analysis

End-to-end OLS regression suite on 11.3M+ transaction rows (75 months, Jan 2020–Mar 2026). ASP confirmed as primary demand driver — price elasticity = −0.901 (R²=27.3%, p<0.0001). Model F (ASP + Seasonality + TET_FEB dummy) achieves R²=0.823, in-sample MAPE=4.68%, out-of-sample MAPE=2.25% on Q1 2026. ARIMA benchmark (SARIMA built from scratch) confirms OLS outperforms pure time-series by 4.4pp — ARIMA cannot observe price changes. Channel elasticities: CVS most price-sensitive (−1.07), NPP near-inelastic (−0.14). UHT most elastic product group (−1.23). Marketing spend (MK** correct codes, ~97B/month) not significant at any lag 0–3 months. All macro variables (CPI, GDP, Rate) formally excluded.

Python OLS Regression ARIMA sklearn scipy 11.3M rows

Live Summary → GitHub

Live

Cost Variance Dashboard (627)

5-page Power BI dashboard analysing manufacturing overhead across 13 factories — YoY variance, cost-per-kg efficiency, and cost-as-%-of-revenue. Built from Oracle EBS GL extracts covering accounts 627, 641, 642 (2023–2025).

Power BI Oracle EBS Power Query DAX VAS

View on GitHub →

WIP

Finished Product Unit Cost — Root Cause Investigation

Monthly root cause analysis of % change in finished product unit cost — decomposing variance into batch data (material consumption, resource usage) and expense data (overhead allocation) to identify and explain cost drivers at the product level.

Power BI Oracle EBS Cost Accounting Variance Analysis

Live

Material Batch Variance Analysis

Full-year 2025 material usage variance analysis across 4 production factories — 27,356 closed batches, 652 finished goods. Joined ERP batch data with INV100 inventory cost (PMAC) to convert quantity variance into VND value. Split RM/SM from packaging material, ranked top 3 best and worst products per factory by value-weighted variance %, and identified key ingredient drivers. Cross-factory packaging analysis reveals systematic yogurt film under-usage across multiple sites.

Python pandas ERP Batch Data Cost Valuation RM/SM Analysis Variance Analysis

Live Page → GitHub

Live

KA Customer Classification — KAC vs KAM

ETL pipeline extracting 15 structured fields from 676 F1 pricing approval PDFs using pdftotext + regex. Built Logistic Regression classifier (CV AUC=0.910) to predict KAC vs KAM customer tier — chosen over Random Forest and Gradient Boosting for interpretability and generalisation robustness on 402 labelled records. Extended to K-Means clustering (K=3,4,5) on 2,219 KA customers from AR ship-to/billing data — K=5 optimal (silhouette=0.168), revealing 5 natural segments: Hospital KAM, Commercial KAC, School & Hotel KAM, Industrial KAM, and National Chain KAC. Key finding: geographic scale (ship-to count, city spread) is the primary KAC differentiator, not industry type alone.

Python PDF ETL Logistic Regression K-Means scikit-learn Unsupervised ML

Live Page → GitHub

Continuous Learning

Coursera Capstone Projects

Hands-on projects completed as part of Google's Advanced Data Analytics certification.

Completed

Employee Turnover Prediction — Salifort Motors

Built and compared 4 classification models (Logistic Regression, Decision Tree, Random Forest, XGBoost) on 14,999 employee records to predict turnover risk. Conducted full EDA, feature engineering (overwork flag, tenure buckets), and model evaluation using precision, recall, F1, AUC, and accuracy. Key finding: employees are systematically overworked — number of projects, tenure, and last evaluation score are the strongest predictors of departure. Logistic Regression achieved 83% accuracy with strong interpretability for stakeholder communication.

Python scikit-learn Logistic Regression Random Forest XGBoost EDA

        Google Advanced Data Analytics Certificate · Course 7 Capstone
      

Live Page → GitHub

Completed

TikTok Claims vs Opinions — Full Analytics Pipeline

End-to-end analytics pipeline across 4 courses: EDA on ~19,000 TikTok videos, hypothesis testing (verified vs unverified accounts, p=2.6×10⁻¹²⁰), logistic regression for verified status prediction, and Random Forest / XGBoost classification achieving ~100% accuracy on claim vs opinion detection. Key finding: engagement metrics alone perfectly predict whether a video makes a claim.

Python EDA Hypothesis Testing Logistic Regression Random Forest XGBoost

        Google Advanced Data Analytics Certificate · Courses 3–6
      

Live Page → GitHub

Where finance
meets data.

Open to the right opportunities.

Where financemeets data.

Open to the right opportunities.

Where finance
meets data.