About

I built a time series forecasting model to predict total daily retail sales using the Favorita Grocery Sales dataset from Kaggle (~3M records across 54 stores and 33 product families, 2013–2017). After cleaning the data — handling nulls in oil price data, removing transferred holiday duplicates, and capping extreme outliers at the 99.9th percentile — I aggregated sales to a single daily total time series and enriched it with two external regressors: a national holiday flag and daily oil price (a known macroeconomic driver in Ecuador's economy).
I trained a Prophet model on ~4.5 years of data, holding out the final 32 days to validate accuracy, then forecasted the following 16-day period. The model achieved a 7.3% MAPE on the holdout set — a strong result for retail forecasting, where 10–20% is typical — and the forecast projected daily sales in the $700K–$1M range for the Aug 16–31 window, capturing weekly seasonality and holiday effects.
Key insight: Sales show clear weekly and holiday-driven patterns, and oil price shows a mild relationship with overall demand. A noted limitation is that the model doesn't explicitly capture Ecuador's well-known "payday" sales spikes around the 1st and 15th of each month — a good next step for improving accuracy further.