Forecasting for decisions

01The problem

Kolanut's Lagos depot orders stock from its suppliers every week, and deliveries take two weeks to arrive. Order too little and the shelves are empty when shops come to restock, so they buy from a competitor. Order too much and money sits in the warehouse, and drinks go past their best-before date.

So every Monday, the depot manager needs a number: how much of each product will sell in the weeks the next order has to cover. Today that number is a guess based on "last month, plus a bit". This course replaces the guess with forecasts built from four years of daily sales, tested honestly, and turned into order quantities.

02The concept

A forecast serves a decision

Before choosing any method, answer four questions:

Time series data

A time series is a measurement taken at regular intervals: daily units here. Unlike the rows in earlier courses, the order matters. Yesterday's sales tell you something about today's, and the past is all you have to learn from.

What forecasts can and can't do

A forecast extends the patterns of the past: trends, seasons, regular events. It can't foresee things that have never happened (a competitor's launch, a strike). Good forecasting is honest about that, with a range as well as a number.

03Example

Load the data. Kolanut's depot sells six products; most of this course follows one of them, the malt drink:

import pandas as pd

base = "https://academy.cloudtechanalytics.com/datasets/demand/"
sales = pd.read_csv(base + "daily_sales.csv", parse_dates=["date"])
holidays = pd.read_csv(base + "holidays.csv", parse_dates=["date"])
print(sales["date"].min().date(), "to", sales["date"].max().date(), "|", sales["product"].nunique(), "products")

malt = sales[sales["product"] == "Malt drink 330ml (24)"].set_index("date")["units"]
malt.resample("YE").sum()

2022-07-01 to 2026-06-30 | 6 products
date
2022-12-31    24519
2023-12-31    48109
2024-12-31    51876
2025-12-31    53298
2026-12-31    24990
Freq: YE-DEC, Name: units, dtype: int64

The first and last years are half-years (July 2022 onwards, and January to June 2026). The full years grow slowly. Now weekly totals for the last few months, the level of detail the ordering decision uses:

malt.resample("W-SUN").sum().tail(8)

date
2026-05-17    1013
2026-05-24     980
2026-05-31     933
2026-06-07     921
2026-06-14     930
2026-06-21     911
2026-06-28     909
2026-07-05     285
Freq: W-SUN, Name: units, dtype: int64

04Walkthrough

1. Load the data and plot the malt drink's daily units: malt.plot(figsize=(12, 4)). You'll see weekly wiggles, a big peak every December, and occasional days at zero.
2. Plot the weekly totals instead. Which patterns are easier to see?
3. Look at holidays: the days the depot was closed explain the zeros.
4. Write down the decision, horizon, granularity and cost of errors for a forecast you know (your own business, or a shop you use).

05Practice

06Check your understanding