Skip to content
Français

Final assignment

  • After the case studies
  • about 3 hr

Where we are

Each module added one piece and tried it alone: the frame, the backtest, the forecasters, the hierarchy, the ranges, the decision and the job. In this assignment, you use them all at once. Larkfield opens a new region and adds 40 products, and you plan the next quarter.

What you will be able to do

Use every module of the course on one new situation, and deliver a forecasting pack that Kenji and Grace can plan the next quarter with. The data looks like the course data, but a new region and 40 new products change it. So you must test each choice again, not copy it.

Before you startWhat to know and to have readyAbout 180 minutes
  • You completed Modules 1 to 8. Each part below names the lessons that it uses.
  • You do not need the case studies of Module 9 for this assignment. (The certificate asks for at least one of them.) To see a similar plan on real data first, open Retail demand with a hierarchy. It cleans, reconciles and orders from a cost table, as you will do here.
  • Your job folder forecast-job works at the end of Module 8, from The forecasting job. If a file is broken, copy the snapshot end-of-m08 from the labs repository again.
  • You can do Parts 1 to 5 in a notebook (local, Colab or Kaggle), with the course’s backtest.py. Part 6 needs the job and a terminal.
  • About 3 hours of work, plus run time. The longest runs in the author’s solution took under 2 minutes each on a laptop.
  • Tested with Python 3.12, pandas 2.3.3, NumPy 2.4.6, statsforecast 2.1.1, mlforecast 1.1.0 and LightGBM 4.7.0 on macOS. The Windows commands are written, not tested.

The situation: a new region and 40 new products

Section titled “The situation: a new region and 40 new products”

Larkfield has grown. On 1 September 2025, it opened a fourth region, east. A fifth region, west, opened on 16 March 2026; it has only two weeks of history, so this plan leaves it out. Since 7 July 2025, it sells 40 new products, P201 to P240, in every region. Now it is Sunday 29 March 2026. The next quarter starts tomorrow: 13 weeks, from Monday 30 March to Sunday 28 June 2026.

Two people need a plan for the whole quarter, today:

  • Kenji, the stock planner, must agree the quarter’s volumes with the suppliers. For them, he splits the quarter into 14-day order cycles, like the orders of the course. That makes seven cycles, and the last one is only 7 days long. He needs one order for each product in each region for each cycle. The weekly job still places the real orders every Monday.
  • Grace, the support lead, must hire temporary agents for the spring. She needs the number of agents for every open hour of the 13 weeks.

Omar, the product manager, asks the question that decides the work: “What will the plan cost us when it is wrong?” Tomás, the AI engineer, has an answer ready: “Use the foundation model for everything. It won the course backtest.” Priya asks you to answer both of them with evidence.

This is the history that you have. Look at it before you read the steps.

  • North
  • South
  • Coast
  • East (from September 2025)
Units sold per week in each region, all products together (the public data, synthetic). The east region starts in September 2025. Every region rises from April to the summer. The quarter that you plan is the next 13 weeks, after the right edge.
Text description of the chart

A line chart with one line per region, 104 weeks from April 2024 to March 2026. North is the largest, between about 2,000 and 5,300 units a week. South is a little lower, and coast is the smallest of the three old regions, between about 1,000 and 2,700. All three rise from April to July, in 2024 and in 2025, and fall in the autumn, with a short peak before Christmas. The east line starts in September 2025, at about 2,100 units a week, and then stays between about 1,400 and 2,400, close to coast. At the right edge, at the end of March 2026, the old regions start to rise again. The east region has no spring yet.

Three things in this chart decide much of the work. The quarter is the spring rise, and the history has seen it only once, in 2025 (the data starts on 1 April 2024). The east region and the new products have never had a spring. And Kenji orders for 13 weeks at once, much further ahead than the 14 days of the course backtest.

All the data is the synthetic Larkfield data. No real people, products or prices.

The data is in the labs repository, in C35/final-assignment/data/. It has the same files and columns as the course data, C35/data/, with the fourth region and the 40 new products. The history runs from Monday 1 April 2024 to Sunday 29 March 2026. The sales of the quarter are not in the files: the course team keeps them back, so that nobody can tune a plan to them.

File Rows What one row is Columns
demand.csv.gz 484,669 One product in one region on one day date, region, product_id, price, on_promotion (1 or 0), sales (units sold), stock_out (1: the shelf was empty, so sales are lower than demand)
products.csv 240 One product product_id, name, category (10 categories), base_price (dollars), launch_date
promotions.csv 458 One planned or past promotion promotion_id, product_id, start, end, discount (0.25 is 25% off)
price_changes.csv 4 One permanent price change product_id, date, new_price, reason
calendar.csv 819 One day, to 28 June 2026 date, weekday, holiday (a name or empty), is_holiday
tickets_hourly.csv.gz 52,416 One help-desk channel in one hour timestamp, channel (chat, email, phone), tickets
tickets_daily.csv 2,184 One channel on one day date, channel, tickets
orders.csv 2,181 One sales channel on one day date, channel, orders
roster_hourly.csv 17,472 One hour timestamp, agents_scheduled
weather.csv 2,912 One region on one day date, region, temp_c (history only)

The folder is about 2.2 MB. SHA256SUMS has a fingerprint of every file, and dataset.md is the dataset card.

What is known in advance: the calendar, the planned promotions and the announced price changes run to 28 June 2026. A shop plans these. The weather of the quarter is not known. Do not use it as an input for the future.

Copy the data next to your job, so that the notebook and the job read the same files. Run this in the folder that holds your clone of the labs repository and your forecast-job folder:

Terminal window
git -C nextia-learning-labs pull
cp -R nextia-learning-labs/C35/final-assignment/data forecast-job/data-final
cd forecast-job/data-final && shasum -a 256 -c SHA256SUMS && cd ../..

On macOS and Linux, each of the ten files must print OK.

The pack follows the course, one module after another. Each part needs the one before it.

Eight boxes in order: the data, 1 frame, 2 read and clean, 3 backtest, 4 hard series, 5 ranges and decisions, 6 the job, 7 the note.
The seven parts of the pack. You choose the methods in Part 3 and test the hard series in Part 4. Part 5 turns the forecasts into the order and the roster. Parts 6 and 7 run and explain the plan.
Text description of the diagram

The diagram has eight boxes in two rows, joined by arrows. First row: the data (two years of sales, four regions, 240 products, tickets per hour), then 1 frame (the order and the roster), 2 read and clean (stock-outs, the missing day, the new region, new products, events), 3 backtest (seasonal naive and three families, with the same cut-offs and measures). Second row: 4 hard series (totals that add up, slow sellers, new products), 5 ranges and decisions (coverage checked, the order and the roster at the cost quantile), 6 the job (a replayed quarter and its bias alarm), 7 the note for Omar.

Make a folder forecast-pack/ in your forecast-job folder. Put every file of the assignment in it. What to submit lists the files.

Uses Too many parasols, too few staff and Frame the forecast.

Write forecast-pack/FRAMES.md, with one frame card for each decision. Use the course’s costs. They are assumptions: write them as assumptions.

  1. The stock order. Write what is forecast (demand, not sales), the level (product and region) and the period (a day). Then write the horizon (the quarter, in 14-day order cycles), how often you make the plan (once, on 29 March), and who uses it.
  2. Write the cost of too low and too high for the order. A lost sale costs 35% of the price (the margin). A left-over item costs 25% of the price (it is sold at a 60% discount). Calculate the critical ratio, the share of the time you want enough stock.
  3. The roster. Write the same parts for the tickets: all channels together, per hour, 13 weeks ahead, the open hours 07:00 to 21:59. One agent answers 6 tickets an hour.
  4. Write the cost of too few and too many agents. An extra agent hour costs 28 dollars. An hour that is one agent short costs 72 dollars: 6 tickets that wait about a day longer, at 12 dollars each. The 72 dollars is an assumption of the course with no measured source. Calculate the critical ratio for the roster.
  5. Add one sentence to each card: the weekly forecasting job still runs every Monday, 14 days ahead. Say how the quarter plan and the weekly job work together.

Part 2: read the data and clean it without hiding demand

Section titled “Part 2: read the data and clean it without hiding demand”

Uses Meet the Larkfield data, Trend, season and noise, Calendars, holidays and events and Clean without hiding demand.

Write forecast-pack/DATA-NOTES.md. It is a table with one row for each problem that you find. Each row says what the problem is, and where it is (product, region, dates). It also says how you found it (the code or the chart), and what you do (fill, mark, model or keep).

  1. Count the rows, the products, the regions and the product and region series. Find the first day of each region and the launch day of the new products.
  2. Find the days with no rows in demand.csv.gz. Fill each one with the course’s rule.
  3. Find the stock-outs with the stock_out column. Count the rows. Fill or mark them, so that a model does not learn an empty shelf as low demand.
  4. Find the events of the quarter in the files that run to 28 June: the holidays, the planned promotions and the price changes. Compare the date of Easter in 2025 and in 2026.
  5. Find the slow sellers: the product and region series with 70% or more days with zero sales.
  6. Write which inputs are known for the quarter and which are not. Give the reason for each.

Uses Naive forecasts, Test like the future: backtesting, Measure the error that matters, Exponential smoothing, Several seasons and long series, Gradient boosting with LightGBM, Deep and foundation models and When not to build a model.

The course backtest forecasts 14 days ahead. Kenji’s plan forecasts 91 days ahead, and it starts at the end of March. So you test three times, and each test answers a different question.

  1. Take the 50 products of the course backtest (BACKTEST_PRODUCTS in the notebooks’ setup cell). Add up their sales in the three old regions, north, south and coast. Fill the missing day.
  2. Run seasonal naive with backtest.py: 13 cut-offs every 28 days, the last one on 15 March 2026, 14 days ahead, the stock cost with the course’s costs.
  3. Compare your row with the course’s backtest table. The old regions and products have the same history as the course data. So your row must be the same: MASE 1.034, WAPE 51.3%, bias −0.5%, cost 16,739 dollars per order cycle. If it is not, fix your data before you go on.
A timeline from April 2024 to June 2026. The history bar runs to 29 March 2026. Ten backtest bars of 91 days start at the cut-offs from 13 April to 21 December 2025. One rehearsal bar starts on 30 March 2025. The quarter you plan starts on 30 March 2026.
The quarter backtest: ten cut-offs, each followed by 91 days. None of them starts at the end of March. The rehearsal from 30 March 2025 forecasts the same 13 weeks one year earlier, the spring rise that the plan faces.
Text description of the diagram

A timeline from April 2024 to June 2026. The first bar is the history, from 1 April 2024 to 29 March 2026. Below it, ten bars of 91 days each: the backtest cut-offs, every 28 days from 13 April 2025 to 21 December 2025. Each bar ends before the end of the history. Below them, one red bar from 31 March 2025 to the end of June 2025: the rehearsal of the same season, one year earlier. The last bar, after the history, is the quarter that you plan, from 30 March to 28 June 2026.

  1. Take the same 50 products, but at the level of the decision: one series for each product in each old region (150 series). Clean them as in Part 2.
  2. Use 10 cut-offs, every 28 days, from 13 April 2025 to 21 December 2025, and a horizon of 91 days. The last cut-off is the latest one with 91 days of real sales after it.
  3. Run seasonal naive as the baseline. Then run at least three other method families. For example: a statistical method (ETS, ARIMA, Theta or MSTL), and a machine learning model (regression or LightGBM with lag and calendar features). Or the foundation model, or a simple rule from the business: last year’s same weeks, scaled to this year’s level. A machine learning model 91 days ahead must not use a lag shorter than 91 days, or it uses values that it will not have.
  4. Give every method the known future inputs that it can use: the promotions, the price and the holidays. Never the weather.
  5. Score every method with the same measures: MASE, WAPE, bias, the fit time and the stock cost per 14-day order cycle. Write the table to forecast-pack/backtest_table.csv.
  6. Do the same for the tickets: all channels per hour, the same 10 cut-offs at 23:00, 2,184 hours (13 weeks) ahead. The baseline is the same hour last week. Score the roster cost per week.

None of the 10 cut-offs starts at the end of March, before the spring rise. A method can win from April to December and still miss the spring. So rehearse the quarter one year earlier.

  1. Before you run it, write your rule for the choice in REPORT.md. For example: “the lowest decision cost on the 10 cut-offs and on the rehearsal together”. Write it first, so that the result cannot change the rule.
  2. Run the same methods from the cut-off of 30 March 2025, 91 days ahead, for the products and the tickets. Some methods cannot run here: the history before this cut-off is only one year long. Say which, and why.
  3. Choose one method for the stock and one for the tickets with your rule. Write the reason in three to five sentences, under the heading Why this method. Answer Omar: what does each method cost in dollars? Answer Tomás: what did the foundation model do in your table, and what would it cost to run?

Uses Hierarchies that must add up, Intermittent demand, New products with no history and Breaks and changes.

Now forecast all 960 series (240 products in 4 regions) for the 91 days from 30 March 2026, with the method from Part 3. Some series need more than that method.

  1. The hierarchy. Larkfield has four levels: the total, the region (4), the region and category (40), and the product in the region (960). Reconcile your forecasts, so that every level is the sum of the level below. Name the method (bottom-up, top-down or MinT) and give the reason. Show with one check that the totals add up and that no forecast is negative.
  2. Slow sellers. For the series with 70% or more zero days, choose a method by the cost, not only by MASE. Remember that “always zero” can win MASE and still lose every sale.
  3. The new region. The east region has no spring in its history. Choose how it borrows a season: for example, the same product in the other regions, scaled to the east’s level of the last weeks. Test the idea: forecast the east’s last 4 weeks from a cut-off 4 weeks earlier. If a test is not possible, give the reason.
  4. The new products. P201 to P240 have about 38 weeks of history and no spring. Choose how they borrow a season, for example the category curve. Test it on a past launch: forecast the new products from six weeks after their launch, as in the lesson.
  5. Events. Put Easter on its 2026 dates, the planned promotions on their days, and the price rise of P051 on 4 May 2026. Say how you estimated the effect of the price rise: the history has other price changes.

Uses A forecast is a range, Score the range and From forecast to order and roster.

  1. Give every forecast a range: at least the 10% and the 90% quantiles of each day, and of each order cycle’s total. If your method has no range of its own, build one from the errors of your backtest.
  2. Check the coverage. Learn the range on the earlier backtest cut-offs. Then count on the later cut-offs how often the real value is inside the 80% range. Do it for the days and for the 14-day totals. Say if the range is too narrow or too wide.
  3. The order. For each product, region and order cycle, order the quantile of the cycle’s demand at the critical ratio from Part 1. Write forecast-pack/orders_quarter.csv.
  4. The roster. For every open hour, plan the agents for the roster’s critical ratio. Plan at least one agent in every open hour, and none outside the open hours. Write forecast-pack/roster_quarter.csv.
  5. The expected cost. Calculate the expected cost of the quarter’s orders and of the roster, from your range. Compare each rule with ordering and staffing the forecast itself, on the later backtest cut-offs. Say which rule is cheaper, and by how much.

Part 6: run the job over a replayed quarter

Section titled “Part 6: run the job over a replayed quarter”

Uses The forecasting job and Watch the forecast.

The quarter plan is made once. The weekly job still plans the next 14 days every Monday, and the bias alarm checks it. Replay the last 13 weeks of the history with the new data, as if the job had run every Monday.

  1. In your forecast-job folder, with its virtual environment active, replay the 13 weeks from 29 December 2025 into a new store:

    Terminal window
    python -m forecast_job replay --data data-final --store runs-final --start 2025-12-29 --weeks 13

    It takes about a minute and a half. Its first line must be the same as the author’s:

    Week 2025-12-29: 22534 units to order (expected cost 50865 dollars), 485 agent hours (expected cost 4911 dollars), 0 warnings, 0 overrides. Plan: runs-final/2025-12-29
  2. Check the last replayed week, which the replay does not check by itself: python -m forecast_job check --data data-final --store runs-final --week 2026-03-23.

  3. Plan the first week of the quarter: python -m forecast_job run --data data-final --store runs-final --week 2026-03-30.

  4. List every bias alarm with its category, its direction and its weeks. For each one, give the likely cause. Use the course’s four answers: the data, a one-time event, a new level, or a pattern that the method cannot learn.

  5. Compare the job’s first 14 days with your quarter plan’s first 14 days, for the total and for the category of each alarm. Say what you would change in the job for the spring, and why. You do not need to change the job’s code.

Uses People and forecasts.

Write forecast-pack/NOTE.md: one page for Omar, Larkfield’s product manager, who does not read code.

  1. Start with the decision: the quarter’s order in units and dollars, the agent hours, and the expected cost of each, with the cost assumptions.
  2. Say how much better the plan is than the simple baseline, in dollars, from your backtest and your rehearsal.
  3. Name the three largest risks, each with the products or weeks that it touches. Include the east region and the new products.
  4. Write what the forecast cannot know: at least the weather of the quarter, and one more thing. Say what the job’s bias alarm will show if it happens, and what Kenji and Grace should do then.
  5. Check every number in the note against your files, as the job’s number check does. If an LLM drafted any part, say so, and check its numbers and its meaning yourself.

The folder forecast-pack/ in your forecast-job folder, with:

forecast-pack/
FRAMES.md Part 1: two frame cards with the costs and the critical ratios
DATA-NOTES.md Part 2: one row for each data problem, and the known and unknown inputs
backtest_table.csv Part 3: one row for each method and test (course check, quarter backtest, rehearsal)
forecast_quarter.csv Parts 4 and 5: ds, region, product_id, yhat, q10, q90 (960 series x 91 days)
totals_quarter.csv Part 4: level, key, ds, yhat (the reconciled totals)
orders_quarter.csv Part 5: cycle_start, days, region, product_id, forecast, low_80, high_80, order
roster_quarter.csv Part 5: hour, tickets, q10, q90, agents (2,184 hours)
replay.txt Part 6: the output of the replay, the check and the first week's run
REPORT.md Parts 2 to 6, and the commands to reproduce every number
NOTE.md Part 7: one page for Omar
notebook or scripts everything that you ran, and requirements.txt with exact versions

REPORT.md must contain these sections:

# Larkfield quarter plan, 30 March to 28 June 2026: forecasting report
## Decisions and costs (short; the full cards are in FRAMES.md)
## Data problems and cleaning (short; the full table is in DATA-NOTES.md)
## Course backtest check (Part 3: your seasonal-naive row against the course's)
## Quarter backtest (Part 3: the table for the products and for the tickets)
## Rule for the choice (Part 3: written before the rehearsal)
## Rehearsal of the season (Part 3: the table, and the methods that could not run)
## Why this method (Part 3: three to five sentences)
## Hard series (Part 4: the hierarchy, slow sellers, the east, new products, events)
## Ranges and coverage (Part 5: coverage on the later cut-offs)
## The order and the roster (Part 5: the rule, the quarter's expected cost, the comparison)
## The job over the replayed quarter (Part 6: the alarms and their causes)
## Limits (what the plan cannot know, and what you would test next)
## How to reproduce (commands, versions, run times, a fresh environment)

There is no reviewer for this course yet. Review your work with this table, or ask a colleague to use it. Mark each item as demonstrated, partially demonstrated or not yet demonstrated, with a reason. The course is complete when every item is demonstrated. You can correct an item and review it again as often as you need.

The points show how much each item weighs, so that you can see where to spend your time. They do not replace the rule: a pack with 90 points and one item not yet demonstrated is not complete.

Required evidence Points Demonstrated when
A forecast frame for each decision 10 Both cards name the target (demand, not sales), the level, the period, the horizon, how often, and who uses it. The costs of too high and too low are given and labelled as assumptions. Both critical ratios are correct.
Data read and cleaned without hiding demand 10 The missing day, the stock-outs, the new region, the new products, Easter, the promotions and the price rise are found, each with where and how. Stock-out days are filled or marked, not trained on as zeros. No future weather is used.
An honest backtest table 20 The seasonal-naive row matches the course’s row. The quarter backtest has seasonal naive and at least three other method families, with the same series, cut-offs, horizon and measures, and the decision cost. The rehearsal of the season is there, and the rule for the choice was written before it. No lag is shorter than the horizon.
Reconciled forecasts and the hard series 15 The totals add up at every level, and no forecast is negative. The slow sellers, the east region and the new products each have a stated method with a reason and a test. Easter, the promotions and the price rise are in the forecast.
Quantile forecasts, the order and the roster 20 Ranges have coverage checked on cut-offs that did not set them. The orders and the roster come from the critical ratios, with at least one agent in every open hour. The quarter’s expected cost is given and compared with the forecast-itself rule.
The job over the replayed quarter 10 The replay output is there, every bias alarm is listed with a cause from the course’s four answers, and the report says what to change in the job for the spring.
The note for Omar 10 One page, no code. The decision and its cost come first. It names what the forecast cannot know and what to do when the alarm fires. Every number is in the pack.
A run that someone else can repeat 5 REPORT.md has the commands, the versions and the run times. The notebook or scripts run from a fresh environment and give the same numbers.

The course does not award completion in four cases, even when every other item is correct:

  • a method is tested with a random split, or with future values in its inputs (a lag shorter than the horizon, the real weather of the quarter);
  • the table has no seasonal-naive baseline;
  • a method is chosen by its error alone, when the decision cost was asked for;
  • the order or the roster is the point forecast, with no range behind it.

A good pack is not the one with the smallest error. It is honest about its measures, it compares every method with a simple baseline, and it ends in a decision with its cost.

These ranges come from the author’s own solution. The author built it from the public data only, with the course’s code. Then the course team scored it once against the quarter that it keeps back. Your methods can differ. Then compare the method and the size of the result, not the exact number.

  • The daily error stays large. At the level of one product in one region on one day, a WAPE of about 60% to 70% is normal, for every method. Most of these series sell a few units a day. Do not expect a method to make the daily error small. Expect it to make the decision cheaper.
  • The totals are much better. The weekly total of all products is the number that Kenji agrees with the suppliers. Its error is about a quarter of the daily product error, or less.
  • A method with no yearly season is too low in spring. In the author’s rehearsal, every method ended about 10% to 16% too low over the quarter. With one year of history, none of them could learn the spring well. Your plan has two springs in its history. On the hidden quarter, the methods with only a weekly season were about a quarter too low on the total. The author’s method, with yearly features, was within about 5%.
  • The best method saves real money. In the author’s quarter backtest and rehearsal, the cheapest method’s orders cost about 20% to 30% less than the seasonal-naive orders. On the hidden quarter, the author’s orders cost about a fifth less than seasonal naive. The roster cost more than a quarter less than “the same hour last week”.
  • The ranges are close to their promise. On the later cut-offs, a good 80% range covers between about 75% and 90% of the days and of the 14-day totals. Outside that, fix the range before you order with it.
  • The ticket method needs the season too. Tickets follow the orders. A ticket model fitted on 12 winter weeks can win from April to December and still miss the spring. The rehearsal shows it.
  • The decision has a number. The note gives the expected cost of the quarter’s order and roster. It also says how much the plan saves against the baseline, in dollars.
  • Choosing the quarter’s method with the 14-day backtest. A method that is good for 14 days can be poor for 91. Test at the horizon of the decision.
  • A backtest with no cut-off in the season. Ten cut-offs from April to December never test a forecast made at the end of March. Rehearse the same season.
  • A lag shorter than the horizon. A model with a lag of 7 days needs its own forecasts to run 91 days ahead. Without them, it uses values that it will not have. Use lags of 91 days and more, or a recursive forecast tested as one.
  • Training on stock-out zeros, or dropping the missing day. The first teaches a model that demand was zero. The second moves every day after the gap to the wrong weekday in a seasonal method.
  • Copying last year’s Easter. Easter was on 20 April 2025 and is on 5 April 2026. Last year’s same weeks put the Easter weekend two weeks late.
  • Copying last year’s promotions. Last year’s promotions are in last year’s sales. Take them out, and put in this year’s planned promotions.
  • A zero forecast for the east and for the new products. They have no last spring, so last year’s same weeks are empty. Borrow a season from similar series.
  • MAPE on series with zero days. It is infinite on a zero day. Use WAPE, MASE and the cost.
  • Totals that do not add up. If the region forecast and the sum of its products are different, Kenji and the suppliers work with two different numbers.
  • The foundation model’s 90% range. Chronos-Bolt gives quantiles from 10% to 90% only. Its “90% range” is its 80% range.
  • Weather as an input. The weather of the quarter is not known on 29 March. A model that uses the real weather looks good in the backtest and cannot run for real.
  • A note with numbers that are not in the plan. Check each number, and read the note for meaning: a number check does not see “orders” written for “units”.
Show the public check values

Open this only after you have your own numbers. These values come from the public data, so you can check them. They are not the answer to the assignment.

  • Data (Part 2): 484,669 rows in demand.csv.gz; 240 products; 4 regions; 960 product and region series. The east region starts on 2025-09-01. P201 to P240 start on 2025-07-07. One day has no rows: 2025-10-19. 3,153 rows are stock-outs (0.65%). The quarter has 76 planned promotions, Good Friday on 2026-04-03, Easter Sunday on 2026-04-05, Easter Monday on 2026-04-06, and one price change: P051 to 11.99 dollars on 2026-05-04. 283 product and region series have 70% or more zero days; 140 of them were on sale from the first day.
  • Course backtest check (Part 3): seasonal naive, 50 products, three old regions, 14 days, 13 cut-offs: MASE 1.034, WAPE 51.3%, bias −0.5%, cost 16,739 dollars per order cycle. With the same data, the 28-day moving average gives MASE 0.875, WAPE 43.5%, and ETS (AutoETS) MASE 0.832, WAPE 41.3%.
  • The replay (Part 6): Replayed 13 weeks from 2025-12-29: 1 bias alarms, 0 weeks stopped. The check of the last week adds one more alarm. The first week of the quarter: Week 2026-03-30: 26065 units to order (expected cost 53497 dollars), 518 agent hours (expected cost 4278 dollars), 0 warnings, 0 overrides.

Recap

You can now take a new situation to a plan that people can act on. You frame each decision by its costs. You read and clean the data without hiding demand. You test methods like the future, at the horizon and in the season of the decision, against a simple baseline. You make the totals add up and give the hard series a method of their own. You turn ranges into an order and a roster with their expected cost. You run the weekly job with its alarm, and explain the plan and its limits to a manager. The case studies of Module 9 use the same steps on real data: Bike-share demand and fleet planning, Electricity load for many customers and Retail demand with a hierarchy.