The dataset that changed my plans had 40,000 rows of mandi price records and a mistake in it. During my economics undergraduate thesis on onion price volatility, I found that the 'weekly average' column I had trusted for two months was computed inconsistently across states — some averaged trading days, others calendar days. Rebuilding it from daily records changed my headline result. What stayed with me was not embarrassment but the realization that the data-cleaning I had treated as clerical work WAS the analysis, and that I wanted the engineering skills to do it properly at scale.
I responded by making my economics degree quantitative rather than switching away from it. I completed coursework in econometrics and statistics, then taught myself Python through progressively less-forgiving projects: first replicating my thesis in pandas, then scraping and reconciling three years of agricultural market data across inconsistent state portals, and finally building a small price-anomaly detector that flagged the 2024 tomato spike two weeks before it made national news — using nothing smarter than seasonal decomposition, honestly applied.
A master's in data science is the deliberate next step because my ceiling right now is method, not motivation. I can fit models; I cannot yet defend the choice of one loss function over another, reason precisely about why my time-series cross-validation leaks, or take a model from a notebook to something a policy team could actually rely on. These are things I could half-learn from tutorials over five years or properly learn from coursework, peers, and supervised research in two.
My direction is applied and specific: agricultural and food-price analytics, where the gap between available data and used data is enormous and the beneficiaries are not advertisers but farmers and food-security planners. An economics graduate who can build pipelines, or an ML engineer who understands markets, is rarer than either alone — I intend to be the overlap.
Every Monday for two years, I have sent a churn dashboard to a leadership team that reads it in ninety seconds. As the business-intelligence analyst at a mid-size lending startup, I know exactly what my dashboards can and cannot say — and the cannot list is why I am applying. When leadership asked which customers would churn next month rather than how many churned last month, I built a logistic regression that performed just well enough to be dangerous: 71% accuracy that everyone treated as prophecy. Watching a retention budget get allocated on my half-understood model taught me the difference between using machine learning and being qualified to.
My work has real analytical substance to build on. I own the SQL layer over our lakehouse, I reduced our reporting latency from daily batch to hourly by redesigning the transformation jobs, and my cohort-level repayment analysis changed how we price one loan product — a decision worth roughly 2% portfolio yield. But every improvement has been within tooling; the statistics underneath are exactly as deep as my last online course, and I feel the difference every time a data scientist joins a call.
I want the graduate coursework that working analysts skip at their peril: statistical inference done properly, optimization, the mathematics under the models I have been calling like APIs. Equally, I want the practicum setting — building under supervision of people who can tell me not just that my model is wrong but the category of wrong. My employer has agreed my role converts to a data-science position on my return, which makes this degree a keyed investment rather than a leap.
Long term, I want to build credit-risk models for the segment my company serves: first-time borrowers with no bureau history, where the standard models fail structurally and the social cost of a bad model is a family denied fair credit. That is a problem worth being properly trained for.