Oman Water Demand Forecasting
Daily 28 · Weekly 11 · Weekly 4

Accuracy = 100 × (1 − WAPE). Every number in this report comes from a real out-of-sample holdout, not from training data. DUQM and PDO are forecast but excluded from the headline score — Chapter 9 explains why, with numbers.

1The question we were asked

A water bottler in Oman sells two products — 24×330ML packs and 30×200ML packs — across 12 sales areas grouped into 7 regions. The factory, the trucks, and the warehouses all need to know how much water will be sold next, and where. We were asked two things:

1. Daily: predict sales for each area for the next 28 days. Trucks load every morning. Target customers: distribution and drivers.
2. Weekly: predict sales for each region for the coming weeks. Production runs, raw water, bottles, and warehouse space are planned weekly. Target customers: plant and supply chain.

2The data we were given

Each row is one day, one area, one product: the day, the area, the region, and the cases sold. Four things about this file shaped every decision that followed.

2.1 There is not a single zero in the file

Across more than 27,000 rows, the smallest sale is 1 case. The source system only writes a row when something was sold. So a missing day usually means no sale happened — not that the data was lost. That single observation drives the filling policy in Chapter 3.

2.2 Some days are missing everywhere at once

20–23 July 2021 (Eid al-Adha) is missing in almost every area. 3 October 2021 (Cyclone Shaheen emergency holiday) is missing in several. Those are real closures, and the pattern repeats across both products — strong evidence they are genuine no-sale days, not broken records.

2.3 Friday means something different in every area

In IBRA, Friday is effectively a closed day. In AMERAT, Friday is a normal selling day. No single "Friday = zero" rule can be true for both. The model has to learn area × weekday, which is exactly what the features in Chapter 4 allow it to do.

2.4 One stretch of October 2021 looks like a data outage, not a holiday

11–15 October 2021: many normally busy 330ML areas (AMERAT, RUSAYL, MUSANNAH, SALALAH, SAHAM) go silent at the same time, while 200ML keeps selling. There is no public holiday in that window. Seventeen values are affected — a tiny number — and they are handled separately from real zero days (Chapter 3), with the assumption flagged in the data.

3How missing days were filled (the policy)

One rule for every blank would be wrong. A blank can mean "nobody ordered that day" (truly zero) or "the system lost that day" (unknown, not zero). Those two need opposite treatment. The policy below sorts every blank into one of four buckets.

The 330ML outage estimate, worked out

Take AMERAT on Monday 11 October 2021. The surrounding Mondays in the source file sold 716, 854, 833, 1732, 690, 912, 837, 653 cases. The median is 835. That 835 is written in, marked as an estimate, and given only a quarter of the training weight of a real row. Seventeen values were filled this exact way.

4How one forecast is made

The model is a small decision-tree ensemble called LightGBM, the same family of method that won the M5 retail forecasting competition. It is not a neural network and not deep learning. We train one model per region per product — ten small models in total in this pack — because each region has its own habits.

4.1 The rule that keeps the score honest

All history features are built with a one-day shift — when the model predicts Tuesday, it is allowed to see Monday and earlier, never Tuesday itself. And the last 28 days of history are held out completely: the model never trains on them. The score in this report is only ever measured on those unseen days.

4.2 What "accuracy" means here (WAPE, in one minute)

Actual week:    1,000 cases
Forecast:         950 cases
Miss:              50 cases
WAPE  = 50 / 1000 = 5%      Accuracy = 100% − 5% = 95%

4.3 The exact settings used (winners only)

Each region's best settings are stored in params/best_params_200ML.json and params/best_params_330ML.json so the reviewer runs exactly what produced these numbers.

Tweedie and Poisson are count-friendly objectives: they suit demand data with many small days and occasional big ones, and cannot predict negative cases. The model stops training automatically when the last 28 days of training stop improving (early stopping).

5Daily results and the daily ceiling

Daily accuracy for the next 28 days, measured on unseen days, five regions:

5.1 Where the daily error comes from

6Weekly results

The same ten models, the same forecasts — simply added up over Monday-to-Sunday weeks before scoring. Two horizons were tested: the last 11 full weeks, and the last 4 weeks.

6.1 Eleven weeks

6.2 The last four weeks

Same models, retrained on history ending four weeks ago — a check that the accuracy is not a lucky artifact of one long window.

7Why weekly beats daily

This is the single most important idea in the pack, and it is not a model trick — it is arithmetic.

Over a week, a slow Monday is usually paid back by a busy Tuesday. Adding the days together before comparing means the random part of the wobble cancels while the real level remains. The model was never bad at the level of demand — it was bad at guessing the exact day. The weekly number uses what the model is actually good at.

8What each forecast is for (business)

9Why DUQM and PDO sit outside the headline

DUQM, DUQUM PDO and PDO are real customers, and the models do forecast them. They are left out of the headline accuracy for one reason: their history is too thin to measure with a percentage fairness.

When a region sells in only a handful of days, a weekly percentage error is divided by a tiny number. Predicting 40 cases in a week when the truth is 10 is reported as a 300% error — a scary number for a 30-case miss. Their absolute misses are small: 200ML PDO's typical daily miss is about 3 cases. That is why the pack reports them separately, with the honest note that a percentage cannot describe them well.

10Reproduce everything yourself

CPU only. No API keys, no internet, no special hardware. From this folder:

pip install -r requirements.txt
python3 daily28/train.py  && python3 daily28/predict.py    # daily 28-day holdout
python3 weekly11/train.py && python3 weekly11/predict.py  # 11-week window
python3 weekly4/train.py  && python3 weekly4/predict.py   # last 4 weeks
python3 make_charts.py                                      # rebuild every chart

11Glossary and straight answers

Straight answers

Why not deep learning? Four years of daily data across a handful of areas is small. Tree ensembles have repeatedly matched or beaten neural networks on this kind of tabular, intermittent demand — including in the M5 competition this project takes its recipe from. Neural networks would add complexity without evidence of gain.

Can daily accuracy ever reach 95%? Not with this data. To get there the model would need to know the events that actually move a single day: promotions and prices, out-of-stock days, delivery schedules, and a confirmed closure calendar. With those inputs a daily model could improve; without them it is guessing at the noisiest part of the signal. Weekly already crosses 95% for four of five regions without any of that.

Does weather help? It was tested properly (temperature, humidity, rain, heat flags from ERA5 reanalysis). It produced no measurable lift on either product, so it was left out to keep the models simple. It can be revisited when more history exists.

What should we do next? Two things pay first: get the promotion and stock-out history from the business to attack the daily error where it is largest, and keep the weekly forecast as the production number, since weekly accuracy is already at the level the plant needs.