Freelance › Projects › Software development › ML Sales & Demand Forecasting Model on Our Historical Data
ML Sales & Demand Forecasting Model on Our Historical Data
Employer
Iryna
Project parameters
Type of cooperationOne-time project
SectionSoftware development
Prepaymentwithout prepayment
Payment methodsCash, Bank transfer
Acceptance of requestsfrom today, 05:28 until Sep 19, 2026
Project description
We run a retail and wholesale business and for years our purchasing and production plans have been built on gut feeling, a couple of spreadsheets and the intuition of two experienced managers. That approach is starting to cost us real money: we either overstock slow items and freeze cash in the warehouse, or we run out of the popular ones right before a spike in orders. We want to move to data-driven planning and are looking for a specialist who can build a machine-learning sales and demand forecasting model on top of the historical data we already have. We can provide roughly three to four years of daily and weekly sales, broken down by SKU, store and channel, plus a calendar of past promotions, price changes and stock-out periods.
The core of the task is a time-series model that predicts demand for the coming weeks and months at the level of individual products and product groups. It has to properly account for seasonality (we have strong yearly and weekly patterns), holidays, and the effect of promotions and discounts, because a naive model that ignores marketing campaigns is useless to us. We expect the work to include honest data cleaning and feature engineering, a comparison of several approaches, and a clear evaluation with accuracy metrics such as MAPE, MAE and bias, reported per category so we understand where the forecast can be trusted and where it still needs a human check. A short written explanation of the results, in plain business language, is just as important to us as the code itself.
Just as important is that the forecast can actually be used every month without a data scientist babysitting it. We need a simple, repeatable way to load fresh sales figures and get an updated forecast — a script or a lightweight interface our own analyst can run, with output in a format we can open in Excel or feed into our planning process. Python is our preferred stack. Please tell us about similar forecasting projects you have delivered, how you would approach seasonality and promotions, and your estimate of realistic accuracy for this kind of data.
— 3–4 years of historical sales by SKU, store and channel
— Seasonality, holidays and promotion effects modelled explicitly
— Accuracy metrics (MAPE, MAE, bias) reported per category
— A monthly re-run process our own analyst can operate
— Clean, documented Python code plus a business-readable summary
The core of the task is a time-series model that predicts demand for the coming weeks and months at the level of individual products and product groups. It has to properly account for seasonality (we have strong yearly and weekly patterns), holidays, and the effect of promotions and discounts, because a naive model that ignores marketing campaigns is useless to us. We expect the work to include honest data cleaning and feature engineering, a comparison of several approaches, and a clear evaluation with accuracy metrics such as MAPE, MAE and bias, reported per category so we understand where the forecast can be trusted and where it still needs a human check. A short written explanation of the results, in plain business language, is just as important to us as the code itself.
Just as important is that the forecast can actually be used every month without a data scientist babysitting it. We need a simple, repeatable way to load fresh sales figures and get an updated forecast — a script or a lightweight interface our own analyst can run, with output in a format we can open in Excel or feed into our planning process. Python is our preferred stack. Please tell us about similar forecasting projects you have delivered, how you would approach seasonality and promotions, and your estimate of realistic accuracy for this kind of data.
— 3–4 years of historical sales by SKU, store and channel
— Seasonality, holidays and promotion effects modelled explicitly
— Accuracy metrics (MAPE, MAE, bias) reported per category
— A monthly re-run process our own analyst can operate
— Clean, documented Python code plus a business-readable summary
