Gains:
- Ability to understand the concepts of time series, trend, seasonality and campaign impact and use artificial intelligence for demand forecasting with the right questions and clean data.
- Ability to interpret forecast accuracy (MAPE, deviation) and uncertainty (confidence interval, scenarios) and base planning on a range rather than a single number
- Being able to distinguish that artificial intelligence prediction is based on historical data and that expert judgment comes to the fore in cases of breaks such as new products, promotions or supply crises.
Every decision in the supply chain is based on one question: how much demand will there be? The more accurately you answer this question, the less stock you hold, the less stock you stay out of, and the less money you tie up. Demand forecasting is the engine of the supply chain; Stock, purchasing, production and transportation plans are all fed from here. Artificial intelligence is extremely powerful in this field because it scans the history of thousands of products much faster than a human and extracts patterns. But the source of its power is also its limit: AI learns from the past; If the future does not resemble the past, it is wrong. In this unit, we will learn the language of prediction, measuring accuracy, managing uncertainty, and where AI should surrender to you.
Language of forecasting: time series, trend, seasonality
Demand data is usually a time series — a series of numbers ordered by time, such as each day's sales of a product. Dividing this sequence into three parts is the basis of prediction:
Trend, long-term direction: are sales increasing or decreasing from year to year? Seasonality, regularly recurring fluctuation: ice cream sells in the summer, the heater sells in the winter; The market is busy on Friday and Saturday. Noise, unexplained random playing. A good forecast captures trend and seasonality and is not fooled by noise. On top of these, events pile up: campaign, promotion, public holiday, weather, competitor move. When you make AI predict, you have to tell it these events; If you don't tell him, he will think that jumping on the campaign day is "normal" and will make a mistake about the future.
Tip: When giving raw sales data to AI, mark campaigns, holidays and out-of-stock days. The day without stock is important: sales appear low that day, but real demand was high, there were just no goods available. If you don't specify this, the AI will underestimate the actual demand.
Measuring accuracy: MAPE and bias
“Is the forecast good?” The answer to the question comes from measurement, not feeling. The most common indicator is MAPE (Mean Absolute Percent Error — how many percentage points on average do predictions deviate from reality). If MAPE is 10%, it means your predictions miss the real value by an average of 10%; lower MAPE is better. The second important concept is bias: is your forecast consistently above (producing overstock) or understocking (leaving understock) reality? Even if a forecast has a low MAPE, if it always deviates in the same direction, there is a systematic problem.
You can give AI part of the history as "training" and part of it as "testing" and compare its prediction with reality. This is the honest way to build trust: testing the model on data it has not seen. A prediction is only worth applying to the future if it has proven consistent in the past.
Caution: No prediction is 100% accurate; It is not expected to happen. The goal is not to produce a perfect single number but a manageable range, knowing the deviation. The binary idea of "the prediction came true/it didn't work" is wrong; The correct question is "how much did it deviate and how much buffer did I keep accordingly?"
Managing uncertainty: range, not single number
The planner's most dangerous habit is to treat the forecast as a single hard number. In reality, every forecast has a confidence interval — like “demand next month is most likely between 900 and 1,100.” The wider this range, the higher the uncertainty. Knowing the range directly affects your stock decision: narrow range requires less safety stock, wide range requires more buffer.
So always ask the AI for scenarios: pessimistic (if demand turns out low), expected and optimistic (if demand turns out high). Seeing three scenarios, the plan is based not on a single assumption but on a range of possible possibilities. This is the most practical way to get out of the "locked in single number" trap.
Step by step: Demand forecasting study with AI
- Clear data. Mark missing days, out-of-stock days, open erroneous records; Work with anonymous product code.
- Tag events. Add campaign, holiday and promotion days to the data.
- Subtract the pattern. Ask the AI to summarize trends, seasonality, and notable events.
- Test it. Store the last part of the history, let the AI predict, compare with reality (MAPE and bias direction).
- Create a scenario. Ask for three predictions, pessimistic/expected/optimistic, and the confidence interval for each.
- Expert filter. AI does not know about known upcoming disruptions (new product, supply crisis, legislation); add them manually. The final plan is yours.
three mini cases
Case 1 — Capturing seasonality. A beverage distributor was complaining that he was constantly out of stock during the summer months and was left with goods in the winter. The planner gave 3 years of anonymous daily sales to YZ and had seasonality removed: June-August demand was 2.4 times the winter average, and the jump started in mid-May each year. According to this pattern, he set up his summer stock 3 weeks in advance; summer dead stock decreased by 70%, winter dead stock fell. AI showed the pattern; The planner made the stock decision.
Case 2 — Giving in to the new product. A company was going to launch a brand new product; There was no historical sales data. The planner asked the AI, "What will be the demand for the first 3 months?" he asked. The AI produced a number that seemed reasonable — but the planner knew it was unfounded: there was no data. Instead, AI asks "What was the first 3-month pattern of products released in similar segments in the past?" analysis and set it up as a forecast range, with cautious starting stock. Expert judgment overruled the AI's fictitious confidence.
Case 3 — Stockless is the trap of the day. An analyst concluded that demand for a product had fallen and recommended reducing inventory. His senior colleague noticed that the low sales days in the data were actually out-of-stock days: demand had not fallen, there were no goods. When the data was corrected with this note and re-given to the AI, it was seen that the real demand increased. Improperly cleaned data produces incorrect predictions.
Four copyable templates
1) Pattern extraction:
Your role: assistant demand planning analyst. Below is the monthly sales data of a product (anonymous code: PRODUCT-X) for the last 36 months. Campaign months [month list], stock-free months [month list]. Task: Summarize (1) the trend, (2) the seasonal pattern, (3) the notable events. Take into account that months without stock hide true demand. The month for which I did not provide data is fabricated.
2) Accuracy test:
Predict the last 6 months using the first 30 of the 36 months below. Then compare my prediction to the actual last 6 months; Calculate the deviation and total MAPE for each month. Tell me if your estimate is consistently biased. Show formula.
3) Scenario forecast:
Estimate next quarter's monthly demand for this product in 3 scenarios: pessimistic / expected / optimistic. Give a confidence interval for each scenario and write down which assumption it is based on. Note that this is not definitive but an interval for planning.
4) Breakage control:
List how the forecast based on historical data I gave you could be invalidated by the following upcoming events: [new product / promotion / supply crisis / regulatory change]. Suggest how I should manually adjust the forecast for each.
Weak prompt / Strong prompt
Weak prompt:
How much will I sell next month?
No data, no product, no context. AI either rejects or makes up; neither result works.
Powerful prompt:
Your role: demand planning assistant. Below are the sales of PRODUCT-X over the last 24 months; Campaign months and stock-free months are marked. Give pessimistic/expected/optimistic forecast for the next 3 months, add confidence interval to each, show MAPE over past test. Pretend I have the final stock decision.
Approach
hit
uncertainty management
Risk
rely on single number
misleading
None
High — plan collapses
Scenario + range
realistic
good
low
just intuition
Variable
weak
medium-high
Blindly obeying AI
Variable
None
high
Common mistakes
- Mistaking a stock-out day for real low demand. When goods are not available, sales appear low; This does not mean that demand has fallen.
- Locking in a single number. To see the prediction as an absolute truth is to ignore the uncertainty; Always ask for range and scenario.
- Not tagging the campaign. The model that thinks promotional bounce is "normal" distorts the future.
- Prediction to new product without history. The seemingly safe number of AI for product without data is unfounded.
- Leaving the breaking to the AI. AI does not know structural breaks such as epidemics, supply crises, and legislation; Expert intervention must be done.
Caution: No matter how good a prediction looks, it cannot be better than the quality of the data on which it is based. "Garbage in, garbage out." Spend most of your time cleaning data; The model part does the rest.
In summary
Demand forecasting is the engine of the supply chain, and AI is a powerful assistant that accelerates this engine. But AI learns from the past; is mistaken when the future does not resemble the past. Capture trend and seasonality, tag events, measure accuracy with MAPE and drift direction, set up forecast as scenario and range, not single number. Expert judgment comes first in new products and structural breaks. Data cleansing is the foundation of everything; The final plan is yours.
Application task
Get sales data for the last 12-24 months of one of your own products (or a hypothetical product), mark the campaign and out-of-stock months. Ask the AI for three scenarios with the “scenario prediction” template. Then store some of the history with the "Accuracy test" template and compare the prediction with the reality. Write down the MAPE you found and the scenario you chose, together with its reason, in 5 items; Also specify which security buffer you will add.
checklist
- [ ] Have I cleared the data, marked out-of-stock and campaign days?
- [ ] Have I validated the prediction with MAPE and bias direction through backtesting?
- [ ] Did I ask for a scenario and confidence interval instead of a single number?
- [ ] If there is a new product / breakage, have I prioritized expert judgment?
- [ ] Have I subjected the final stock plan to my own approval?