New products with no history
- Lesson 20
- 35 min
- Module 6: Hard series and many series
Where we are
We can forecast slow sellers. A new product is harder still: it has only a few weeks of history, or none. In this lesson, we borrow from similar products, the category curve and a global model, and decide how much to trust the first weeks.
Lesson 20 of 27 · Forecasting and Time Series: Predict What Comes NextNew products with no history
1. Forecasts serve decisions
2. Read a time series
3. Baselines and honest testing
4. Statistical forecasters
5. Machine learning forecasters
6. Hard series and many series
7. Uncertainty and the decision
8. Run a forecast in production
What you will be able to do
Forecast a product that has only six weeks of sales. Its own history is short and starts slowly after the launch. You borrow from what you know: its own last days, its category’s shape and a global model trained on all products. Then you measure how far you can trust each one.
NotebookDo this lesson in a notebook
The notebook has every code cell, task and check of this lesson. It builds the category curve by hand, forecasts Larkfield’s 19 new products with four methods, and scores each product on its own. Its setup cells install the packages and download the Larkfield data, so it needs no local setup. About 35 minutes, of which about 20 seconds is waiting for LightGBM.
Colab and Kaggle open your own copy. They need a free Google or Kaggle account. Stuck? See the solution notebook.
Before you startWhat to know and to have ready
What you need to know
- WAPE and bias, from Measure the error that matters.
- A global LightGBM model with lags, from Gradient boosting with LightGBM.
- The seasonal naive forecast (the same weekday last week), from Naive forecasts.
What you need on your computer
Nothing: the notebook runs in Colab or Kaggle. Its first cell also installs mlforecast and LightGBM 4.7.0.
Tested with Python 3.12, pandas 2.3.3 and LightGBM 4.7.0, in Jupyter on macOS.
A new café in town
Section titled “A new café in town”A friend opens a café. After six weeks, she asks you: “How many croissants should I bake next week?”
Her own six weeks help, but they are not normal weeks. In the first week, few people knew the café. Sales grew each week. You could also ask another café in town: “What does your week look like?” Its Saturdays are busy and its Tuesdays are quiet. Your friend’s café is smaller, but it probably has the same shape.
So you combine two things: the size from her own weeks, and the shape from a similar café. That is the main idea of this lesson.
The problem: six weeks that do not look like the future
Section titled “The problem: six weeks that do not look like the future”Larkfield launched 19 products after the first day of its data, with at least six weeks and 14 more days of sales. Kenji plans the first full order on the product’s 41st day after launch. Priya tests that moment for all 19.
The garden hammock P007 is an example. It was launched on Monday 7 April 2025. Its first week sold 13 units in total. Its sixth week sold 287 units:
- Units sold per day
- Category curve (WAPE 18.3%)
- Own mean of 14 days (22.7%)
- LightGBM global (32.4%)
Text description of the chart
The chart shows the hammock’s daily sales from its launch on 7 April 2025. In the first two weeks, it sells 0 to 5 a day, with one day of 13. Sales grow to between 26 and 65 a day in the sixth week. For the next 14 days, from 19 May to 1 June, the real sales are dots between 16 and 48 a day: 481 in total. The category curve forecasts 477, the mean of its last 14 days 456 and the global LightGBM model 382. The seasonal naive forecast, not drawn, repeats the sixth week, with days of 65 and 59: 574.
The short history has three problems:
- It starts slowly. The launch weeks are lower than the normal level, so an average of them is too low.
- It is short. Six weeks hold six Mondays. A weekly pattern from six values is noisy.
- It has no year. You cannot know how a hammock sells in autumn from its spring.
The idea: borrow from similar products
Section titled “The idea: borrow from similar products”A forecast for a new product borrows information from older products. Three common ways:
- Own recent mean: use only the newest part of its own history, the last 14 days. It skips the slow start, but it has no shape.
- Category curve: take the shape from the average old product in the same category, and scale it to the new product’s size. Shape from the neighbour, size from yourself.
- Global model: one model trained on all products, like the LightGBM of Gradient boosting with LightGBM. It sees the new product’s age, category, price and promotions, and learns from every product how a product of that age and kind sells.
A similar method is analogue products. You choose by hand a few old products like the new one (same price, same use), and copy their launch curve. Larkfield’s data starts on 1 April 2024, so we cannot see the launch curves of the older products. Here the category stands in for the analogues.
The category curve by hand
Section titled “The category curve by hand”A new lamp sold 30 units in its first 2 weeks. The average old lamp of its category sold 20 in the same 2 weeks. So the new lamp is 30 ÷ 20 = 1.5 times an average lamp. Last week, the average old lamp sold 2, 2, 3, 3, 4, 6 and 5 units, Monday to Sunday.
| Mon | Tue | Wed | Thu | Fri | Sat | Sun | |
|---|---|---|---|---|---|---|---|
| Average old lamp, last week | 2 | 2 | 3 | 3 | 4 | 6 | 5 |
| × 1.5 = new lamp, next week | 3 | 3 | 4.5 | 4.5 | 6 | 9 | 7.5 |
ratio: 1.5 | next week: [3.0, 3.0, 4.5, 4.5, 6.0, 9.0, 7.5]The same steps with the hammock’s real numbers:
Text description of the diagram
The diagram has three steps. Step 1, how big is the new product: the hammock sold 647 units in its first 6 weeks; the average old garden-furniture product sold 267.9 in the same 6 weeks; the ratio is 647 ÷ 267.9 = 2.415. Step 2, what shape the category has: the average old product’s sales last week, Monday to Sunday, per day: 11.87, 11.87, 12.65, 13.78, 14.17, 18.82 and 15.61. Step 3, forecast equals shape times size: each day times 2.415, for 2 weeks: 28.7, 28.7, 30.6, 33.3, 34.2, 45.5 and 37.7. In 14 days, that is 477 units; the hammock sold 481.
Larkfield: four methods on 19 new products
Section titled “Larkfield: four methods on 19 new products”Priya forecasts all 19 new products at their 41st day, with four methods. The fourth is one global LightGBM model. Its inputs are the age of the product in days, the weekday, the category, the price and the promotion flag. It also gets the sales of 14, 21 and 28 days before. It learns again at each product’s origin, from all 200 products.
own mean WAPE % and bias %: (35.7, 2.9)seasonal naive WAPE % and bias %: (47.9, 10.8)category curve WAPE % and bias %: (35.2, -12.7)LightGBM global WAPE % and bias %: (45.3, 6.6)Text description of the chart
A column chart of the WAPE of four methods on the 14 days after the 41st day of 19 new products. Own mean of the last 14 days: 35.7%, bias +2.9%. Seasonal naive: 47.9%, bias +10.8%. Category curve: 35.2%, bias −12.7%. LightGBM global: 45.3%, bias +6.6%.
Three results stand out.
- The two simple borrowing methods win, almost tied: the category curve 35.2% and the own mean 35.7%.
- The category curve is 12.7% too low. Its ratio uses all six weeks, including the slow launch weeks. So it underestimates a product that has reached its normal level. The own mean of the last 14 days skips the start, and its bias is only +2.9%.
- The global model loses. In Module 5, LightGBM was among the best methods on the 50 old products. Here it is 10 points behind. It learned mostly from products with two years of history, where the lags carry most of the information. A new product’s lags are few and low, and its first weeks have none.
How much can you trust six weeks?
Section titled “How much can you trust six weeks?”A total WAPE hides the spread between products. Here is each product’s WAPE, one dot per product:
Text description of the chart
A dot plot with four rows, one per method, and 19 dots per row, one per new product, on an axis from 0 to 200%. Own mean of the last 14 days: from 21.4% (gloves P097) to 113.0% (parasol base P027). Seasonal naive: from 27.4% to 175.0%. Category curve: from 18.3% (hammock P007) to 101.0% (lawn seed P069). LightGBM global: from 22.3% to 191.3%. In every row, a few products are near or above 100%: the parasol base P027, the lawn seed P069 and the hedge trimmer P099. The best method per product is the own mean for 8 products, the category curve for 8, the seasonal naive forecast for 2 and LightGBM for 1.
The answer to “how much can you trust six weeks?” is: not much for a single product. The hammock’s category curve missed 18.3% of the units. The parasol base P027 missed 95% or more with every method. It sold 241 units in its first six weeks, in the spring, and only 46 in the next two weeks. The lawn seed P069 and the hedge trimmer P099 sold so little (27 and 22 units in six weeks) that every forecast was mostly noise.
So Kenji treats a new product’s forecast differently:
- Order for a shorter time, for example one week instead of two, and look again.
- Keep a wider range. Module 7 turns a forecast into a range; for a new product, that range should be wide.
- Watch the bias of the first orders. If the first two orders are too low, the product has probably grown past its launch weeks.
When to use each, and what it costs
Section titled “When to use each, and what it costs”| Method | Use it when | It fails when | Cost |
|---|---|---|---|
| Own recent mean | the product has at least 2 calm weeks | the product is still growing fast, or has a weekly pattern | nothing |
| Category curve | the category has a clear shape and similar products | the product is unlike its category (a parasol base sells unlike parasols) | a few lines of pandas |
| Global model | many new products, with features that explain them (price, promotions, age) | the new products are few, and the model learned from old ones | one model per origin (19 here, about 20 seconds) |
| Analogue products | an expert can name similar launches | nobody can, or the history before the launch is lost | a person’s time |
With no history at all (before the launch), only the category curve with a planned size and the analogues remain. The size then comes from Kenji’s plan: “we expect it to sell like a top-third garden-furniture product.”
Common mistake
The launch weeks in the average
Symptom: Priya uses the category curve with a ratio over all six weeks for every new product. Kenji’s first orders run short: over the 19 products, the forecasts are 12.7% too low.
Cause: the ratio counts the launch weeks, when few customers knew the product. The hammock sold 13 units in its first week and 287 in its sixth.
Fix: check the bias, not only the WAPE. In the notebook, a ratio from the last 14 days changes the bias to +9.4%. But the WAPE gets worse: 38.1% against 35.2%. Two weeks of one new product are few and noisy. A middle way, such as the last four weeks, or the own mean for products that grow fast, is worth a backtest.
Consequence in AI work: a method with the best total error can still be wrong in one direction. In a stock order, a forecast that is always too low means empty shelves for every new launch.
Guided practice: new products in the notebook
Section titled “Guided practice: new products in the notebook”- Open the notebook from the box Do this lesson in a notebook at the top of this page.
- Run the setup cells. The data cell ends with the
Ready:line. - In section 1, predict the lamp’s Saturday. Then run the cell. You see
9.0for Saturday. - In section 2, run the cell. You see
19 new products. - In section 3, predict if the category curve is too high or too low. Run the cell. You see
-12.7. - In section 4, predict if LightGBM beats the category curve. Run the cell. You see
(45.3, 6.6). - In section 5, run the cell. You see each product’s WAPE and the best method per product.
- In section 6, write your answer in the Your turn cell, then run the check cell.
Your turn: a ratio from the last two weeks
Section titled “Your turn: a ratio from the last two weeks”Before you run it, predict: if the ratio uses only the last 14 days, will the WAPE go down, and will the bias change sign? In the notebook, section 6, compute the category curve with a 14-day ratio for the 19 products, and put its WAPE % in wape_recent.
Show the answer
pairs = []for p in new.itertuples(): mine = sales[sales["unique_id"] == p.product_id] hist = window(mine, p.origin - pd.Timedelta(days=14), p.origin)["y"].to_numpy() real = window(mine, p.origin, p.origin + pd.Timedelta(days=14))["y"].to_numpy() others = sales[(sales["category"] == p.category) & sales["unique_id"].isin(old_ids)] cat_2_weeks = window(others, p.origin - pd.Timedelta(days=14), p.origin).groupby("ds")["y"].mean() cat_last_week = window(others, p.origin - pd.Timedelta(days=7), p.origin).groupby("ds")["y"].mean().to_numpy() pairs.append((real, np.tile(cat_last_week, 2) * hist.sum() / cat_2_weeks.sum()))print("category curve, 14-day ratio, WAPE % and bias %:", wape_bias(pairs))wape_recent = wape_bias(pairs)[0]category curve, 14-day ratio, WAPE % and bias %: (38.1, 9.4)The bias changes sign: from −12.7% to +9.4%. The WAPE gets worse: 38.1% against 35.2%. Removing the slow start fixed one problem and added noise. Keep a change only when the WAPE and the bias both say so.
Knowledge check
Section titled “Knowledge check”Check your understanding
Explain the category curve with the hammock's numbers, and say why it was too low on average over the 19 new products.
Show the answer
The hammock sold 647 units in its first six weeks. The average old garden-furniture product sold 267.9 in the same weeks, so the hammock is 2.415 times an average product. The category’s last week, per day (11.87 to 18.82), times 2.415 gives the forecast. That is 477 units in 14 days, against 481 sold. Over the 19 products, the category curve was 12.7% too low. Its ratio includes the slow launch weeks, when few customers knew the product.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
Tomás says: 'LightGBM was among the best models in Module 5, so it must be best for new products too.' What did Larkfield's backtest show, and why?
Show the answer
On the 19 new products, the global LightGBM model had a WAPE of 45.3%. The category curve had 35.2%, and the own 14-day mean 35.7%. It was best for only 1 of the 19 products. The model learned mostly from old products with two years of history. For them, the lags of 14, 21 and 28 days carry most of the information. A new product’s lags are few, low and partly missing, so the model has little to work with.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
Predict: Larkfield launches 40 new products next spring. Kenji asks how far he can trust their first forecasts. What would you tell him, with numbers?
Show the answer
For a single product, not far. In the backtest, the WAPE of single new products ran from about 20% to 100% or more for every method. The parasol base P027 missed 95% or more with all four. Over many products, the errors partly cancel: the best methods had a total WAPE of about 35%. So I would suggest shorter orders for new products, a wide range, and a check of the bias after the first two orders. A spring launch, like the parasol base, may also fall fast after its season.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Recap
- A new product’s first weeks are short, start slowly and have no year.
- Borrow: the own recent mean skips the slow start. The category curve takes the shape from the category and the size from the product. A global model learns from all products.
- On 19 new products at six weeks: category curve 35.2%, own 14-day mean 35.7%, LightGBM 45.3%, seasonal naive 47.9%.
- The category curve was 12.7% too low: its ratio counts the launch weeks. A 14-day ratio flipped the bias to +9.4% and made the WAPE worse (38.1%).
- Trust six weeks little for one product: single-product WAPE ran from about 20% to above 100%. Order for shorter times and watch the bias.
New products change from week to week. Old products can change too, all at once: a price rise moves their level for good. Next, you find such a break and test whether the forecast should forget the old history.
Next: Breaks and changes