15. Machine Learning for Demand Forecasting

Teaching rewrite — the Baseline model step. Original wording; illustrative numbers. Competition figures recomputed from the published scores.

Introduction

Statistical models (lesson 14) apply predefined equations: give a non-seasonal model a seasonal product and it mistakes the cycle for changing trends; give a seasonal model a flat product and it invents seasonality. Machine learning (ML) works the other way round: a learning algorithm discovers relationships between inputs (history, drivers) and outputs (future demand) from data.

This lesson covers:

• What ML is — and isn’t — and how it learns from features
• Feature engineering as a business activity
• Local vs global models
• Black box vs white box, and how to peek inside (feature importance, scenarios)
• Tree-based models and neural networks, with a short history
• Realistic expectations from forecasting competitions and practice
• How to launch an ML initiative — and four pitfalls

---

15.1 What is machine learning?

flowchart LR
  subgraph Statistics
    D1[Data] --> M1[Predefined model] --> F1[Forecast]
  end
  subgraph Machine learning
    D2[Data] --> L[Learning algorithm] --> M2[Learned model] --> F2[Forecast]
  end

You choose the algorithm and its settings, but not the resulting relationships — the model learns them from a training dataset, then applies them to new data.

	Statistical models	Machine learning
Explainability	White box	Black box
Complexity	Simpler	More complicated
Demand drivers	Hard to include	Easy to include
Data requirement	Limited	Extensive
Expected accuracy	Lower	Higher

Neither replaces the other — they trade transparency for flexibility.

What counts as ML here

Tree-based models (random forests, gradient-boosted trees) and neural networks. Linear regressions (including Lasso/Ridge/ElasticNet), ARIMA, and Prophet are treated as statistical models — be wary of vendors relabelling them as ML. Be even warier of “AI” claims: the term has no strict definition and often sells over-promises.

Features

A feature is a piece of information the model can use to predict, e.g. “last quarter’s demand”, “next quarter’s price”, “promotion next quarter?”, “brand”.

Brand	Demand Q−4	Q−3	Q−2	Q−1	Avg price	Price Q+1	Promo Q+1	Demand Q+1 (target)
Low cost	1,500	500	400	300	10	10	—	200
Premium	500	1,000	750	500	35	30	−15%	700
Regular	250	350	150	400	20	20	−20%	300

(This is supervised learning: the model sees inputs and the desired output.)

Two success factors:

1. The features you provide — business insight matters most. You can’t predict Black Friday sales without knowing whether the item will be discounted.
2. Tuning the algorithm — a data-science skill.

Feature engineering is a business process

Don’t leave data scientists alone. Run a brainstorm: “If I had to forecast next month myself, what would I want to check?” Typical answers become features:

• Current price, and did it change recently?
• Average monthly demand?
• Was it recently out of stock?
• Is a promotion running?

Then the hard part: collecting the data. Trade off quality vs availability — don’t freeze a project for months to collect a barely useful driver.

Local vs global models

	Local (typical statistical)	Global (typical ML)
Trained on	One item’s history at a time	All items at once
More history helps?	Yes	Yes
More items helps?	No	Yes — more examples of each pattern
Transfer learning	None	A never-promoted item can benefit from what promotions did to other items
Small datasets	Works	Struggles — too few examples

A global model learns general patterns (“promotions lift demand”, “recent growth tends to continue”); it learns cross-item relationships only if you give it features describing them.

Opening the black box

ML won’t tell you level, trend, or seasonality. Two tools help:

• Feature importance (tree models) — shows which features the model relies on, not how they affect demand. Useful to drop useless drivers (weather at 0.01% importance). If a driver you consider critical scores low, suspect bad data or a design flaw.
• Scenario analysis — run the model with and without a driver (e.g. with/without a discount) and compare. Example: a promotion scenario shows +17% total uplift and dips before/after the promotion (self-cannibalisation).

---

Checkpoint: ML fundamentals

Quiz

Checkpoint — section 15.1

1 / 4

What distinguishes ML from statistical forecasting models?

Pick an answer to see if you are right.

---

15.2 Main types of learning algorithms

Technical extra — skip to 15.3 if you only need the business view.

A short history

Year	Milestone
1940s	McCulloch & Pitts model artificial neurons
1957	Rosenblatt’s Perceptron (first trainable neural network); Holt’s exponential smoothing
1960	Winters adds seasonality to exponential smoothing
1963	Morgan & Sonquist: first decision tree algorithm
1985	Damped trends for exponential smoothing (Gardner & McKenzie)
1986	Backpropagation popularised — deeper networks trainable
1995–2001	Boosting (AdaBoost 1995/97), random forests (2001)
2010–2015	ReLU activation, Adam optimiser — deep networks train well
2016–2017	XGBoost, LightGBM — fast, accurate gradient boosting

Why the boom only recently? Computing power and data grew, but the decisive factors were better algorithms (2010–2017) and simpler tools (open-source libraries, learning resources).

Tree-based models

A decision tree asks yes/no questions (nodes) until it reaches a prediction (leaf) — like the game Guess Who?. The training algorithm picks, at each node, the question that splits the data into the most homogeneous groups. (“Does the person wear clothes?” is useless — everyone does.)

Example path for a forecast:

flowchart TD
  A{Demand Q−1 > 300?} -- yes --> B{Average price > 20?}
  A -- no --> X[Predict 150]
  B -- no --> C{Price Q+1 < 25?}
  B -- yes --> Y[Predict 650]
  C -- no --> Z[Predict 550]
  C -- yes --> W[Predict 420]

• Random forest — averages hundreds of independent, randomised trees; simple and robust.
• Gradient boosting — builds trees one after another, each focusing on the previous trees’ mistakes; today’s top performers (XGBoost, LightGBM).

Neural networks

An artificial neuron weights its inputs and applies an activation function (often something like a capped weighted sum, e.g. ReLU) to produce an output.

flowchart LR
  I1[Demand Q−2] --> H1((h))
  I2[Demand Q−1] --> H1
  I3[Discount] --> H2((h))
  I4[Price] --> H2
  I1 --> H2
  I4 --> H1
  H1 --> O1[Forecast P1]
  H1 --> O2[Forecast P2]
  H2 --> O1
  H2 --> O2

• Input layer passes features in; hidden layers do the learning; output layer produces forecasts (one neuron per period).
• Training = an optimiser adjusts the connection weights to reduce error.

---

15.3 What to expect from ML

Forecasting competitions

Competition	Year	Granularity	Items	Horizon	Benchmark	Benchmark score	Winner	Error reduction
Corporación Favorita	2018	Store × product × day	210,000	16 days	Seasonal naïve	0.8486	0.5092	40%
M5 (Walmart)	2020	Store × product × day	42,840	28 days	Exponential smoothing	0.6710	0.5204	22%
Intermarché	2021	Store × product × day	275,781	90 days	Seasonal moving average	0.6088	0.5478	10%

Error reduction = (benchmark − winner) / benchmark. All three were won by ML. Gains differ partly because of the benchmark (a weak seasonal naïve flatters Favorita) and the data/metric (Intermarché had one year of history and a log-scaled metric). Each used a different metric, so compare cautiously.

Rules of thumb from practice

• ML typically reduces forecast error by 0–30% versus a good moving average.
• Gains grow with more data — shortages, promotions, prices; promotions alone can be worth up to ~15%.
• ML shines most on granular data.
• Claims far beyond this deserve scrutiny — check for cherry-picked periods or excluded “worst” items.

Improving the baseline, not replacing planners

ML may not beat your whole process (planners know things the model doesn’t), but it should beat your baseline — freeing planners to focus on new products and items where they have special information.

Step	Before: MAE (FVA)	Bias	After ML: MAE (FVA)	Bias
Benchmark	52%	−1%	52%	−1%
Baseline model	45% (+7)	−3%	41% (+11)	−2%
Demand planners	41% (+4)	+1%	39% (+2)	+1%

The final forecast improves from 41% to 39%, and planners need to add less — less workload.

---

Checkpoint: algorithms and expectations

Quiz

Checkpoint — sections 15.2 and 15.3

1 / 4

How does gradient boosting differ from a random forest?

Pick an answer to see if you are right.

---

15.4 Launching an ML initiative

Follow the 5-step framework (objective → data → metrics → baseline model → review). For ML, the model-building loop gains a step:

flowchart LR
  A[Visual data analysis] --> B[Model selection] --> C[Feature selection] --> D[Parameter optimisation] --> E[Error analysis] --> A

Four pitfalls

Pitfall	Avoid it by
Bad test set-up	Hold out a test period never used for feature selection or tuning (e.g. tune on 2021–2025, test on 2026). Picking the best of many models on the same period you report is cherry-picking. Compare against a benchmark, your current engine, and your consensus — same periods, same granularity
“What” before “why”	Don’t chase fashionable models. Fix the objective (level, horizon), the data (demand, not sales), and the metrics (MAE + bias, not MAPE) first
Wrong expectations	Promise too much → disappointment and user resistance; too little → no traction. Plan for 0–30%; >15% is realistic when promotions or pricing drive demand
“We tried ML once”	ML is hundreds of models and set-ups. One failed attempt doesn’t prove ML can’t work — any more than one failed spreadsheet proves maths doesn’t

---

Rank: ML project

Quiz

ML project sequence

1 / 2

Rank the steps of an ML forecasting project.

OrderDrag the handle to rearrange — top should be first.

1. ⋮⋮1Agree metrics (MAE + bias, value-weighted) and benchmarks.
2. ⋮⋮2Collect demand data and key drivers (promotions, prices, shortages).
3. ⋮⋮3Select features and tune models on a training period.
4. ⋮⋮4Evaluate on an untouched test period against benchmark, current engine, and consensus.
5. ⋮⋮5Deploy as the baseline; track FVA.
6. ⋮⋮6Define the decision-level objective (granularity, horizon).

Drag items into the correct order (or use ↑↓), then check.

---

Check: machine learning

Quiz

Machine learning for demand forecasting

1 / 5

Which model is a typical example of ML in this lesson?

Pick an answer to see if you are right.

---

Takeaways

Remember

• ML learns relationships from data; statistical models assume them
• Features drive success — make feature engineering a business exercise
• Global models learn across items; need enough data
• ML is a black box — use feature importance and scenarios
• Expect 0–30% error reduction vs a good benchmark; more data (promotions, prices, shortages) → bigger gains
• ML improves the baseline and reduces planner workload; it doesn’t replace planners
• Test honestly on held-out periods; follow the 5-step framework

Further Reading

• 14. Statistical forecasting
• 12. Forecast value added
• Next: 16. Judgmental forecasting
• S. Makridakis et al., “The M5 accuracy competition: Results, findings and conclusions” (2022)
• Background: N. Vandeput, Demand Forecasting Best Practices (Manning) — used as a study outline