Joint Machine Learning Competition with Panasonic Group

Tokyo Electron (TEL) held a joint machine learning (ML) competition with the Panasonic Group in November 2025. The goal was to enhance employees’ machine learning skills and promote the exchange of technical talent. This article is an event report contributed by the TEL ML competition organizing team.

Background:
One Email Sparked a Joint Competition

TEL has been running ML competitions for employee skill development for some time. The idea of a joint event started around January 2024, when a TEL employee saw a blog article on a Panasonic Group ML competition and reached out to use their dataset for TEL’s internal competition. After several exchanges about each other’s internal competitions, the idea of a joint competition took shape. It took about a year to organize and carry out.

What is an ML Competition?

An ML competition is an event where participants develop machine learning models to solve a given task using provided data, and then compete on the performance of these models. Specifically, it often involves supervised learning to achieve the highest predictive accuracy.

For this competition, we used a publicly available dataset modified to focus on a demand forecasting task. Mr. Sakata from the Panasonic Group, who holds the highest Kaggle rank of “Grandmaster,” created the task. Kaggle is a well-known platform for ML competitions and learning.

This was the 8th internal competition at TEL, with dozens of participants each time. This time, the Panasonic Group joined, bringing the total number of participants to 141. The event was held over two weeks from November 5 to 19, 2025, and was very successful.

Practical Task Setting:
It’s Not Just About Improving Prediction Accuracy!

The data centered on typical demand forecasting. The task aimed to minimize loss caused by discrepancies between predicted and actual demand, assuming predictions would be used directly in production planning. If demand is predicted accurately, sales will increase – conversely, inaccurate predictions can lead to losses. Because the products varied and the level of prediction criticality differed by product, simply maximizing prediction accuracy was not enough. This setup closely reflected real-world business challenges.

Operations and Results

Due to the companies’ information security rules, it was difficult to share the same environment. So, each company ran the competition in their own system, exchanged leaderboard data, and combined results to create a joint ranking.

On the first day, Mr. Sakata appeared as a guest in the tutorial session. He carefully explained exploratory data analysis (EDA), a naive benchmark mode^*1 without ML, feature engineering^*2, and the workflow for building models. Both companies held the opening and closing ceremonies together.

A total of 141 participants submitted at least one model, with 4,416 submissions overall.
The final results are shown in the table below.

While the results include technical terms from machine learning, the first-place participant used LightGBM^*3 finely tuned quantiles, and diverse features. The runner-up combined multiple time-series models^*4 such as Chronos and TimesFM. Many top participants applied quantile regression^*5 showing how well the task suited methods that leverage the data and problem characteristics.

1. Naive benchmark: A simple model used as a baseline to evaluate new methods.
2. Feature engineering: Creating new features from raw data to boost model performance.
3. LightGBM: A model combining many simple decision rules, developed by Microsoft.
4. Time-series-based model: A deep learning model created by learning a large amount of time-series data.
5. Quantile regression: Predicting specific points in data distribution beyond the average, helpful for understanding data variability.

Reflections on the First Joint Competition and Future Goals

This first joint competition received many positive comments such as “the joint format increased motivation” and “Mr. Sakata’s tutorial was very educational.” It became a fun opportunity to improve machine learning skills. We are truly grateful to the Panasonic Group for their cooperation.

We will continue to encourage friendly competition and skill development together. TEL aims to contribute to enhancing corporate value and will keep dedicating efforts to human resource development activities.