Comparative analysis of ARIMA and Prophet using Amazon India sales data — 128,975 transactions, validated against the Global Superstore dataset, resolving exactly when each model wins.
The research was designed around a single operating question for e-commerce leaders: which forecasting model should we trust, and when?
Does Prophet's AI-driven seasonality decomposition outperform ARIMA's classical statistics for e-commerce demand — and does the answer change with portfolio size?
Prophet wins at 1–2 categories (30.74% lower MAE). ARIMA wins at 3+ categories. A precise "3-category tipping point" governs model choice.
A tested decision framework that cuts forecast error by up to 32.76%, reduces inventory waste, and lowers the carbon footprint of logistics operations.
ARIMA remains the industry default for stable, linear demand — but e-commerce is volatile, seasonal, and fast-moving. Prophet was purpose-built for exactly that pattern, yet practitioners have no data-driven rule for choosing between them.
Inaccurate model selection compounds into overstock, higher warehousing cost, product wastage, and a larger carbon footprint — the study set out to replace guesswork with a tested framework.
Implement ARIMA (1,1,1) and Prophet on 128,975 real Amazon India transactions using Python in Google Colab.
Evaluate both models with MAE and RMSE at category-level and total-demand level, then identify the category-count "tipping point."
Validate findings on the independent Global Superstore dataset and quantify the sustainability impact of optimal model choice.
Quantitative, descriptive-comparative design. Secondary data, positivist paradigm, fully reproducible in Google Colab (Python 3.12).
128,975 raw transactions, 24 variables, Kaggle
Cancelled orders & nulls removed → 110,643 rows
Daily Qty by category, chronological split
77 days train / 14 days test
ARIMA(1,1,1) & Prophet (weekly seasonality)
MAE & RMSE via scikit-learn
Aggregate demand across the full nine-category Amazon India portfolio, March–June 2022.
14-day holdout test for the highest-volume "Set" category (42,947 records). Hover any point for the exact value.
Weekly seasonality captures day-of-week shopping rhythm, keeping error 30.74% below ARIMA at single-category granularity.
Predictions converge near the 77-day mean (~430 units/day), reflecting limited autocorrelation in volatile category-level data.
Four operational read-outs drawn directly from the category-level and aggregate-level results.
At 1–2 product categories, demand carries strong weekly seasonality. Prophet's trend + seasonality decomposition captures this directly, cutting MAE by up to 30.74%.
Once three or more categories are aggregated, individual volatilities cancel out. The smoother series favours ARIMA's linear trend modelling — a ~32% error reduction.
A single "one model fits all" policy is measurably suboptimal. Portfolio size — not intuition — should decide which engine drives replenishment planning.
Run Prophet at SKU/category granularity for top sellers; run ARIMA at aggregate warehouse level. The hybrid approach captured the best of both in this study.
Estimated impact of switching to the category-optimal model across the full portfolio.
Estimated annual inventory cost savings at 9-category total-demand optimisation.
Maximum forecast error reduction achieved at multi-category aggregate level.
Peak inventory units saved per day through optimal model selection.
Reduced forecast error translates directly into less overstock, fewer emergency shipments, and a smaller logistics carbon footprint.
Enter your own order volume and warehousing cost to estimate savings from switching to the category-optimal forecasting model — using the actual waste-reduction percentages measured in this study.
The identical pipeline was re-run on the independent Global Superstore dataset (2011–2014, 51,290 records, 3 categories).
128,975 records · Set category · Prophet MAE 37.47 vs ARIMA 54.10
51,290 records · Office Supplies · Prophet MAE 30.56 vs ARIMA 51.37
Prophet wins at single-category level on both datasets — 40.51% advantage confirmed.
A consulting-format summary of the full research, condensed for executive sign-off.
Answer three questions about your own forecasting setup and get a data-backed recommendation, grounded in this study's actual results.
Select a category count to see live MAE / RMSE comparisons and the recommended model update in real time.
In e-commerce logistics analytics, a category is a product grouping used to aggregate SKU-level sales into a single demand time-series — e.g. "Fashion > Sets," "Fashion > Kurtas," or "Electronics." Warehouses and forecasting systems rarely predict demand item-by-item; instead they forecast at the category level, since that's the level at which inventory, shelf space, and replenishment decisions are actually made.
In this project, each category count on the slider below represents how many of these product categories were summed together into one daily demand series before forecasting — starting from a single category ("Set" alone, 42,947 records) and scaling up to all 9 categories in the Amazon Sale Report dataset combined. This is exactly what drives the result: forecasting one narrow, volatile category behaves very differently from forecasting nine categories pooled into one smooth aggregate signal — which is why Prophet wins at low category counts and ARIMA overtakes it at three or more.
Hover over any bar to preview its numbers live; click a chip to lock that category count in.
Seven chapters — literature review, methodology, full data analysis, conclusions, and recommendations — as submitted for the MBA (Artificial Intelligence & Data Science) capstone.
Puvvada Rohan Sai Pavan · EA2452001011709 · SRM Institute of Science and Technology · May 2026
MBA — Artificial Intelligence & Data Science · SRM Institute of Science and Technology, Kattankulathur · Academic Year 2024–2026
Guided by Dr. Veena Christy, Assistant Professor. Independent research conducted end-to-end in Python (Google Colab) using publicly available Kaggle datasets.