A demand forecasting calendar to solve the problems of "overproduction" and "shortage" in make-to-stock production.

Introduction — Common challenges of make-to-stock production

"How many more should we produce for next month's sale?"

Food, cosmetics, consumer goods, parts—all companies that manufacture to make-to-stock face this question.

According to a survey by the Ministry of Economy, Trade and Industry, lost sales and waste due to inventory management errors in the manufacturing industry amount to several percent of sales. Despite this, many workplaces still rely heavily on the experience and intuition of veteran employees when deciding how to increase production.

For companies where e-commerce sales are becoming the main source of revenue, the situation is even more serious. With Rakuten, Amazon, Yahoo!, and Furusato Nozei (hometown tax donation) programs, more than 30 sales events occur annually. Managing all of this mentally is simply impossible.

In this column, we will introduce the functions and technology of the demand forecasting calendar built on ERPNext, specifically for business leaders.

---

1. What you can do with this feature

---

🗓️ 1-1. Dashboard for centralized management of business information

The demand forecast calendar is not just a forecasting tool, but a company-wide dashboard that aggregates information essential for business management.

The main components are the following three tabs:

Tab	Content	Business Benefits
🔥 Demand Events	Monthly calendar of e-commerce sales and seasonal events	See at a glance when and what will happen
🏭 Production	Daily Production Results vs. Plan	Instantly check if production is progressing according to plan
📈 Prediction Accuracy	AI prediction accuracy displayed by item	Confirm prediction reliability numerically

In addition, it also features tabs for sales/orders, accounting, and purchasing/inventory, allowing you to get a comprehensive overview of data from each module of ERPNext.

▼ This is a screenshot of the actual demand forecast calendar interface.

---

🤖 1-2. AI-powered automated demand forecasting

Behind the scenes, Prophet, an open-source time-series prediction AI developed by Meta (formerly Facebook), is running. Three prediction engines are available, which can be used depending on the application.

Engine	Forecast Period	Main Applications
Preliminary Notice Forecast	8 weeks ahead	Manufacturing plan development
EC Forecast	2 weeks ahead	Decision on replenishing inventory
Company-wide forecast	3 months ahead	Management decisions and financial planning

Our preliminary forecasts are used for manufacturing planning, our e-commerce forecasts for inventory replenishment, and our company-wide forecasts for management decision-making—each providing forecasts with the optimal level of detail and timeframe.

---

🛒 1-3. Automatically calculate the impact of e-commerce sales.

The system automatically calculates the effective impact of each event captured. The key point is to "avoid uniformly increasing production." Instead, it allocates the impact based on the actual sales channel composition ratio over the past 90 days.

Example: During the Rakuten Super Sale (adjustment rate +30%), if Rakuten accounts for 73% of sales, the effective impact is +21.9%. This can be used as a guideline for actual production increases.

If we uniformly increase production by 30%, we'll end up producing too much for channels other than Rakuten, leading to waste and losses. This system automatically eliminates that risk.

---

⏪ 1-4. Lead Time Calculation and Alert Notifications

The system automatically calculates the lead time backward from the event start date and sends incremental alert notifications.

Example: Regarding the start of the sale on March 4th

• February 20th — Deadline for ordering raw materials (procurement lead time: 7 days)
• February 27th — Manufacturing start deadline (manufacturing lead time 3 days)
• March 2nd — Deadline for preparing shipments (shipping lead time: 2 days)

You will receive an alert bell 🔔 three days before each deadline, so in this example, you will be reminded gradually starting 12 days before the event. Lead time settings can be customized to suit your company's specific needs.

---

🛡️ 1-5. Visualization of prediction accuracy and safety measures

"Are AI predictions really accurate?"

To answer this question, the MAPE (Mean Absolute Percent Error) for each item is calculated in real time and displayed in a list with a signal indicator (🟢 within 20% / 🟡 within 30% / 🔴 needs improvement).

Furthermore, we have also put in place safety measures in case our predictions are wrong.

Solution	Content
Manual Override	Individual values ​​can be manually changed. Changes will not be overwritten when the AI ​​is rerun.
Batch Percentage Adjustment	Apply a ±N% adjustment to specific items and periods in a single application.
Automatic Retuning	Parameters are automatically readjusted weekly when accuracy exceeds a threshold.

---

2. Overview of the Technology

---

🧠 2-1. Why Prophet is suitable for forecasting production make-to-stock

The reasons why Prophet is suitable for make-to-order manufacturing are clear.

Characteristics	Meaning in Manufacturing
Automatically recognizes annual and weekly seasonality	Automatically learns patterns such as "Frequent shipments on Fridays" and "Increased shipments towards the end of the year"
Can incorporate the impact of holidays and events	Japanese holidays and e-commerce sales can be added as special days
Multiplicative Seasonality	If sales double, the fluctuation range also doubles—a model that reflects reality
Lightweight operation without GPU	No cloud AI required. Works perfectly well on your own servers.

Although Meta archived the Prophet repository in 2023, it is still functioning without issues. Its successor, NeuralProphet, is API compatible, making future replacements low-cost. The prediction logic of this system is consolidated into a single method, so the impact of library replacement will be limited.

---

🔄 2-2. Autonomous Improvement Loop

Typically, machine learning models are periodically tuned by data scientists, but in this system, AI automatically measures accuracy and readjusts parameters.

Step	Processing Details
① Accuracy Measurement	Compare predictions and actual results for the past 2-4 weeks to calculate MAPE
② Threshold Determination	Tuning begins when MAPE exceeds 30%
③ Parameter Search	Try up to 48 combinations of key parameters
④ Save optimal values ​​	Automatically adopt the most accurate parameter set

In addition, we automatically calculate daily adjustment rates based on past manufacturing performance (for example, reflecting the practice of manufacturing more on Fridays to include Monday's shipments), and apply post-production adjustments.

No data scientists are needed. The system automatically learns and improves continuously.

---

🔐 2-3. Security — Data does not go outside.

Sales data and customer information are our most important trade secrets. This system is designed to be completely self-contained on our own servers.

• Sales, order, and forecast data → Stored in the company's ERPNext DB.
• AI model → Trains and performs inference using our own CPU (no external APIs used)
• ICS hub → localhost communication only. No external transmission.
• Access control → Apply ERPNext's role-based access control.

We have zero reliance on cloud AI services.

---

⚠️ 2-4. The Limitations of AI — I'll be honest

Pattern	Reason	Solution
New Product	No historical data available	Display "Insufficient Data" and skip prediction
SNS Buzz	Unpredictable based on past trends	Manually adjust by ±N%
Disaster/Pandemic	Exceeds AI detection range	Partial response through automatic adjustment of trend sensitivity parameters
Sudden termination of business with a trading partner	No zero assumptions	Manual override

These limitations can all be addressed through manual overrides.

---

3. Installation Guide

📋 Prerequisites

Conditions	Details
ERPNext v15 or later	No issues if using the ERPNext.JP environment
Past Sales Data	Sales Invoices for 30 days or more
EC orders must be stored in ERPNext	We can also discuss EC integration.

📅 Startup Schedule

Timeframe	Milestones
Week 1–2	Custom app implemented. Sales events automatically generated. Calendar operation started.
Week 5–6	Data accumulation enables company-wide forecasting
Week 9–12	Monitoring on the accuracy dashboard. Initial automatic tuning.
90 days ~	Autonomous tuning fully operational. Accuracy is stable.

💰 Cost

No additional server costs or external cloud services are required. It runs on your existing ERPNext server. Daily operations are fully automated, and it can be operated with only a monthly accuracy review (approximately 15 minutes).

---

4. Summary — Reinforcing "Experience and Intuition" with "Data and AI"

This system does not negate on-the-job experience. It is a mechanism for quantifying experience and accumulating it as organizational knowledge.

Traditional Operation	After Implementation
Manually manage sale schedules	Automatically add to calendar
Determining increased production volume based on experience	Automatically calculating effective impact based on past performance
You realize it at the last minute and panic	Gradual alerts starting 12 days in advance
Concerns about AI accuracy	Constant monitoring on the dashboard
Experts are needed for adjustment	AI self-tunes
Concerns about data leaks	Completely within our own servers

---

Technology Stack

Elements	Technology
Foundation ERP	ERPNext (Python / Frappe Framework)
Prediction Engine	Prophet (Open Source Time Series AI)
Event Aggregation	ICS Hub (Microservices)
External Cloud	Not required (Completely handled on our own servers)

---

If you have questions like, "Can you make predictions with our products?" or "How much data do we need?", please contact us. 30-minute free online consultation Please feel free to ask us anything. You can start using the calendar function in as little as two weeks.