0
Skip to Content
OSCAR ANALYTICS
Oscar Analytics
PROJECTS
OSCAR ANALYTICS
Oscar Analytics
PROJECTS
Oscar Analytics
PROJECTS

How Long Until $1M? A Bootstrap Simulation

Techniques: Monte Carlo simulation, bootstrap resampling, reproducible random-seed design — built in Python (NumPy, pandas, Matplotlib).

A Monte Carlo simulation projecting the path to a $1,000,000 net worth under real market uncertainty, built around a personal financial model (income, savings, a Roth IRA, debt paydown, and market-exposed investing) run 1,000 times over.

Each simulation bootstrap-resamples actual historical S&P 500 annual returns (1961–2024) rather than assuming one constant average return, so the output is a genuine distribution of possible outcomes rather than a single misleadingly precise number, visualized as 1,000 overlapping net-worth trajectories over time.

A Surrogate Model for Net Worth Forecasting

Techniques: Metamodeling, regression, random forest, feature importance analysis — built in Python (scikit-learn, pandas, NumPy, Matplotlib).

A machine learning model trained to replicate my Monte Carlo net-worth simulation instantly, without re-running it.

Thousands of synthetic scenarios were generated by varying inputs like salary, rent, debt, and contribution rate, then a random forest learned the relationship between those inputs and time-to-$1,000,000.

Feature importance analysis reveals which financial levers matter most, and predictions were validated against fresh simulation runs to confirm the surrogate holds up.

SORTING THE NEWS: SEMANTIC SEARCH AND CLUSTERING

Techniques: Sentence embeddings, unsupervised clustering (KMeans), dimensionality reduction (UMAP), cosine-similarity search, cluster validation (Adjusted Rand Index) - built in Python (sentence-transformers, scikit-learn, pandas, Matplotlib)

A random sample of 3,000 AG News articles across four categories - World, Sports, Business and Sci/Tech - encoded with a pre-trained sentence-transformer into 384-dimensional vectors, so that stories with similar meaning land close together in a single embedding space.

KMeans groups the articles without even seeing the category labels, and an Adjusted Rand Index of 0.636 against the true categories measures how much of the real structure it recovers, so the result is a quantified agreement rather than an eyeballed one – visualized as side-by-side UMAP projections of the true and discovered groups. The same embeddings power a cosine-similarity search that ranks articles by meaning rather than exact keywords.

doES A ONE-PAGE CHECKOUT SELL MORE? AN END-TO-END A/B TEST

Techniques: SQL, power analysis, sample-ratio-mismatch detection, CUPED variance reduction, Bayesian inference, bootstrap resampling, A/A testing

Analyzed a 60,000-user experiment comparing a one-page checkout with a three-step checkout, from a pre-registered plan to a ship decision. I found bot traffic unbalancing the test groups, measured a +17% lift in conversion, and caught a novelty effect that would have overstated the long-run gain by about 2.5×. The data is simulated with a hidden answer key, so each method was checked against the true effect.

oscar analytics

info@oscar-analytics.com