GitHubLinkedIn

Featured work

Complete project stories,
without leaving the site.

Each project opens as a structured report with the question, analytical decisions, verified findings and honest limitations.

BITCOIN NLP / 20226 TOPICS · 3 FORECASTS
Largest topic share59.9%Crypto / Ethereum discussion
Crypto / Ethereum59.9%
Expressive language19.0%
Check-mark patterns9.1%
Index / pointing6.0%
Lightning network3.6%
Promotional cluster2.3%
−0.153 Baseline test R²−0.184 Sentiment test R²Finding: no reliable forecast gain
01 / Research

Bitcoin NLP × Volatility

An end-to-end natural-language-processing study that maps Bitcoin discussion themes, measures daily sentiment and tests whether social signals improve next-day volatility forecasts.

View report →
USER RETENTION MODELEXPLORATORY
17.7%observed churn

Champion model XGBoost

Accuracy81.6%
Precision45.1%
Recall18.3%
F126.1%
14,299 labelled users3-way train / validation / testLow recall limits deployment
02 / Modelling

Navigation-App Churn

A full churn-analysis case study using behavioural data, statistical testing, feature engineering and interpretable model comparison to assess which users may disengage.

View report →
NYC TAXI / FARE MODELLEAKAGE CAVEAT
Reported test R²0.868Re-estimation required after leakage fix
Test MAE$2.13
Test RMSE$3.79
Trips analysed22,699
p < .001 Fare / payment associationResults shown with limitations intact
03 / Modelling

NYC Taxi Analytics

A multi-stage analytics project that moves from trip-level exploration and hypothesis testing to fare regression and tree-based classification on New York City taxi data.

View report →
DATA QUALITY REPORTIN BUILD
92/100Illustrative quality score
Completeness94%
Consistency88%
Validity96%
01 Upload02 Inspect03 Clean04 Export
04 / Application

Clean My Data

A Streamlit application that identifies common data-quality problems, supports cleaning decisions and produces a clear exploratory overview of uploaded datasets.

View report →
CANONICAL CORRELATIONCOMPUTER FEAR × STATISTICS FEAR
First canonical r0.74856.0% shared variance
F1
F2
F3
F4
F5
F6
F7
2,571 respondents4 / 7 significant functionsOne dominant multivariate relationship
05 / Research

Fear Patterns with CCA

A canonical correlation analysis of how a seven-item computer-fear profile relates to an eight-item statistics-fear profile across 2,571 respondents.

View report →
CLAIM FREQUENCY / FREMTP L2VALIDATED
Zero-claim policies96.32%677,991 policies audited

Final test comparison

XGBoost Poisson0.2355 deviance0.3446 Gini
Nonlinear Poisson GLM0.2416 deviance0.2918 Gini
Aggregate A/E1.00008XGBoost final refit
26,444 observed claims0.0738 claims / policy-yearMore ranking power, calibration retained
06 / Modelling

Insurance Claim Counts

An exposure-aware comparison of Poisson-family GLMs and XGBoost for insurance claim frequency, with validation-led selection, DHARMa diagnostics, calibration and explainability.

View report →
BTC-USD / 2015-2026VALIDATION SELECTED
GARCH-LSTM validation QLIKE1.9425Lowest pre-test loss
Test QLIKE1.7304
Test RMSE13.58
Daily returns4,239
Robustness5 seeds · 3 folds
GARCH · 82%
LSTM · 18%
p = .0026 Hybrid vs standalone GARCHSmall neural gain, strong econometric core
07 / Modelling

Bitcoin Volatility Forecasting

A validation-led study comparing econometric, machine-learning, neural-network and hybrid models for daily Bitcoin variance forecasting from 2015 to 2026.

View report →
NHANES / MISSING DATAMICE · m = 10
Records retained21,187Pre- and Post-COVID observations
IncomePMM
BMINORM
CholesterolNORM
SmokerLOGREG
BMI0.07 → 0.51
Cholesterol−2.75 → 0.01
~13% overall missingness20 iterationsPooled uncertainty leads the conclusion
08 / Research

Multiple Imputation

A missing-data study using NHANES observations, chained equations and pooled uncertainty to compare BMI and cholesterol across Pre- and Post-COVID periods.

View report →
AUTO MPG / BAYESIAN NETWORKPROBABILISTIC MODEL
In-sample accuracy58.7%392 vehicle records
Expert graph8directed edges
Learned graph5after pruning
Agreement3common edges
DisplacementWeight
DisplacementHorsepower
HorsepowerAcceleration
26.65% Largest posterior at weight 3,0006 MPG categoriesInference with uncertainty visible
09 / Modelling

Bayesian Network MPG

A Bayesian-network study comparing an expert-defined vehicle dependency graph with a BIC-learned structure, then using conditional probability tables for MPG inference and classification.

View report →
Browse featured projects

Upcoming projects

Ideas taking shape.

Projects currently being planned and developed for Toni's Data-Works.

Idea in development

Extreme value theory and financial risk

Does Diversification Survive the Tail?

EVT-GARCH backtesting of Value at Risk and Expected Shortfall for a ZAR-denominated multi-asset portfolio.

Question being explored

Do EVT-based risk models estimate the frequency and severity of extreme losses more reliably than historical simulation and conventional GARCH?

Historical simulationStudent-t GARCHPOT-GPD EVTFiltered EVT-GARCHVaR backtestingExpected Shortfall

Planned application: rolling risk forecasts for South African equities, US equities, gold and Bitcoin, with losses translated into rand.

In development

Big data, actuarial analytics and forecasting

Mortality Modelling Across Two Countries

A comparative study of Czechia and Slovakia using deep learning and interpretable statistical modelling.

Question being explored

Can a deep neural network improve age- and sex-specific mortality forecasts beyond a generalised additive model while remaining stable on future years?

Deep neural networkGeneralised additive modelChronological validationMortality forecastingMAE / MSE / R²

Inspired by Richman and Wüthrich's neural-network mortality framework, the project uses Czechia and Slovakia as a paired setting for comparing predictive flexibility with model interpretability.

Continuous learning

Certifications

Training that supports the work shown in this portfolio.
  1. 01

    Google Advanced Data Analytics Professional Certificate

    Google · Coursera

    Statistics, regression, machine learning, Python and analytical communication

    2026
  2. 02

    CHPC-NITheCS Coding Summer School

    CHPC · NITheCS

    Scientific computing, programming and computational problem-solving

    2026
  3. 03

    Google Data Analytics Professional Certificate

    Google · Coursera

    Data preparation, SQL, R, Tableau and evidence-based decision-making

    2025
  4. 04

    Investment Banking: Financial Analysis and Valuation

    University of Illinois Urbana-Champaign · Coursera

    Financial statements, company analysis and valuation fundamentals

    2025
  5. 05

    Citi Investment Banking Job Simulation

    Citi · Forage

    Investment-banking analysis and transaction-oriented problem-solving

    2025
  6. 06

    Standard Bank Behavioural Economics Job Simulation

    Standard Bank · Forage

    Behavioural insights, customer decisions and practical intervention design

    2025
© 2026 Toni Catatonidataworks