This guide walks through the data, the baselines you must beat, the features, the models, the error measures that work when sales are often zero, and the scope for each level.
- Data: the M5 competition set from Walmart has 42,840 time series across three US states, with a 28-day forecasting task.
- Beat a baseline first. A forecast that cannot beat “same day last week” has no value.
- Split by time, never at random. The test period must come after the training period.
- Lag and calendar features with a gradient-boosted model are a strong and explainable approach.
- Avoid MAPE when products often sell zero. Use MAE, RMSE and a scaled error.
Which dataset should you use for demand forecasting?
The M5 Forecasting data is real retail sales supplied by Walmart. The organisers’ own account of the competition is in Makridakis, Spiliotis and Assimakopoulos (2022), in the International Journal of Forecasting.
| What | Detail | For your project |
|---|---|---|
| Source | Daily unit sales from Walmart stores | Real data, with zero-sales days and promotions |
| Series | 42,840, from single items up to departments, categories and stores | Pick one store or one category to start |
| Where | California, Texas and Wisconsin | Compare one store per state for a larger project |
| Task | Point forecasts 28 days ahead | Hold out the last 28 days as your test period |
Figures are from the competition organisers. A local shop’s own sales records also make a good, original dataset if the owner agrees to share them.
How do you build a demand forecast step by step?
- Choose a slice you can explain: one store, one category, a few hundred items.
- Split by time. Train on the past, validate on the next 28 days, test on the final 28.
- Set the baselines: last value, same weekday last week, and a moving average.
- Build features: sales lagged by 7, 14 and 28 days, rolling means, day of week, month, price and event flags.
- Fit a classical model, such as exponential smoothing or Prophet, on the total series.
- Fit a gradient-boosted model such as LightGBM on the feature table, one model for all items.
- Compare every model on the same test period, per item and in total.
- Build the dashboard: actual against forecast for any item, the error, and a reorder quantity.
Which error measures should a forecasting project report?
- MAE: the average miss in units, easy for a shop owner to read.
- RMSE: punishes large misses more.
- MASE or another scaled error: your error divided by the naive forecast’s error. Below 1 means you beat the baseline.
- Bias: whether you forecast too high or too low on average, since the two mistakes cost differently.
Leave MAPE out when items often sell zero units, because dividing by zero sales breaks it.
What mistakes cost marks in a forecasting project?
- A random train and test split. It lets the model see the future.
- Lag features that peek ahead. For a 28-day forecast, a lag of 1 day is not available on day 20.
- No baseline. An error figure means nothing without something to compare it with.
- One chart of one good item. Show the spread of errors across items.
- Fitting scalers or encoders on all the data before the split.
How does the scope change for Diploma, B.Tech and M.Tech?
| Level | Scope | What to show |
|---|---|---|
| Diploma | Weekly sales of a few products: a moving average against linear regression | A chart of actual against forecast, and next week’s order |
| B.Tech / B.E. | One M5 store: baselines, Prophet and LightGBM compared on a held-out 28 days | A Streamlit dashboard with a search by item |
| M.Tech / M.E. | A base paper reproduced, plus one extension: forecasts that add up across the hierarchy, prediction intervals, or deep models against boosted trees | A significance test across items and a paper in IEEE format |
A delivered M.Tech project of ours is also a forecasting system: network traffic prediction with a Bi-LSTM and a live dashboard. More time-series topics are among the deep learning ideas beyond images.
What goes in the report, and which viva questions come up?
Start with the stocking problem, then the data, the validation scheme, the baselines, the models, the results and what the shop should order. Our report and PPT help covers the black book and slides. Expect these:
- Why did you split by time instead of at random?
- What is a seasonal naive forecast, and did you beat it?
- Which lag features mattered most?
- Why is MAPE a poor choice here?
- How would you forecast a new product with no history?
Other tabular guides: customer churn prediction and credit card fraud detection. Polytechnic students can start from our Diploma projects page.
This guide is a plan to discuss with your guide, and the figures in it come from the linked dataset pages and papers, not from our own work. Our 8 delivered case studies are M.Tech and M.E. projects, each shown with its real paper pages, screens and results.
Frequently asked questions
Which dataset is best for a demand forecasting project?
The M5 Forecasting data on Kaggle is the best known. Walmart supplied it for the M5 competition: 42,840 time series of unit sales across stores in California, Texas and Wisconsin, with a task of forecasting 28 days ahead. Use one store or one category; the full set is heavy for a laptop.
Which algorithm is best for demand forecasting?
It depends on the data. Start with a seasonal naive forecast and exponential smoothing. Then try a gradient-boosted model such as LightGBM on lag and calendar features, which is a popular choice for retail data. Deep models such as an LSTM are worth adding only if they beat these.
Is demand forecasting a good final year project?
Yes. It solves a problem every business has, the data is public and the result is easy to show on a dashboard. It also teaches time-based validation, which many classification projects never touch, and examiners notice when it is done properly.
Why should I not use MAPE for demand forecasting?
MAPE divides the error by the actual sales, and many products sell zero units on many days, so the figure becomes undefined or huge. Use MAE or RMSE, and a scaled error such as MASE that compares your forecast with a naive one.
Can The Ultimate Project World help with a demand forecasting project?
Yes. Share your level, your review dates and any data you already have in the free consultation. We will tell you exactly which parts we can take on, such as the model, the dashboard, the report, the PPT and viva preparation.
