Pricing Analytics with AI — Part 2

Ozan Çevik
KoçDigital
Published in
4 min readMay 25, 2023

In the first post of this series, I wrote about how recession and inflation affect businesses and how we as KoçDigital offer help with decision making via our analytics services in such turbulent times.

Our next topic of interest will be “Can we automate Procurement with Autonomous or Semi-Autonomous Negotiations?”, but first I would like to introduce this new project of mine, which needs combining lots of external data, and mainly focuses on insights driven price deals and negotiation process.

Spare Parts Purchasing Price Prediction Case

Problem Statement:

Recently, I started this interesting price prediction project in which the customer is a leading global automobile manufacturer working with various suppliers. To achieve a more organized supplier operations, a dedicated team is assigned to each combination or category of suppliers. These teams are responsible for making purchase orders and price negotiations with vendors on a chosen portal.

The project deals with quite an elaborate set of external data (both open source and paid) varying from some go-to financials to sector-specific raw material prices. While internal know-how of the employees for this data set is extremely deep, but data contains some inaccuracies/errors and extreme outliers which are hard to define and detect even for this team. As a part of the project, we strenghthen the customer’s external data set by adding some financial conversions on certain currencies and future currency predictions which is included in our DaaS, Data as a Service product.

The company wants to convince vendors with lower and more realistic prices empowered by strong AI-based algorithms and mathematically strong heuristics. Also, another significant goal of the company is to lower the annual budget with targeting to record positive price deals per order. Finally, for the company, combining these results via dashboards and systems integration to explain price predictions both for vendors and purchasing responsibles easily is another critical component of the project. In summary with this project company aims; (i) decision support activities related to negotiations with vendors (min %75–80 accuracy for all products); (ii) support budgeting and procurement planning activities; (iii) improve the performance of the annual supply chain budget with AI contribution around %10 reductions.

Methodology:

Since we have worked with 10 years of data, we applied common data cleaning algorithms and outlier detection models. All outliers and wrong data entries are marked/seperated with the consent of subject matter experts of customer teams. Our initial parts’ clustering methodology failed due to SAP ERP based master data BOM (Bills of Materials) structure inefficiency.

We used four machine learning algorithms which are LightGBM, XGboost, Elastic Net Regression and Artificial Neural Networks. For the accuracy metric we have selected weighted MAPE(wMAPE).

While searching academics and literature, I realised that AI based price prediction research focuses onSupply Chain field is limited. While there are different and valuable studies on electricity or commodity prices, for a pure automotive supply chain pricing problem I could only find Alalawin et al’s article. Also, demand prediction approaches are highly sophisticated and crowded in literature, however price prediction for finished products are lacking.

In “Forecasting vehicle’s spare parts price and demand” the research team concentrates on using Artificial Neural Network algorithms in the price and demand predictions in electric cars.

In our project data quality and failure in clustering force us to find more advanced techniques since our customer was asking to prove the external data relationship we applied correlation coefficient for all materials and all external data sources. After getting %80 +/ and -%80 correlation coefficients with the external factors we asked our customer to label each factor for the legitimacy of the causality whether those results are logical or a coincidence.

After receiving confirmation we have applied 4 core algorithms mentioned previously. We applied price segmented clustering (0–1,1–5,5–20…) for materials that do not have strong correlation coefficient and automatic pattern finding for external data.

Explaining those might be a little bit confusing, but the figures below represent the models fit clearly.

3 Models that were run with 4 algorithms; LGBM,Elastic Net,ANN and XGBoost.

Results:

· Our results cover February-March 2023 price predictions and 12 months of past data which are used to test the accuracy of the results with completed deals.

· We have converted all data set in US dollars to make valid predictions.

Scores on average 12 months can be seen below.

Even in demand forecasting projects those results are extremely hard to achieve. Pricing prediction with a large data set of 30–40 k lines is much more harder to guess. %84 accuracy for 12 months rolling is a strong validation.

Since we have quite accurate guesses, the company decided to show these predictions on the purchasing screen. Purchasers will decide whether to use AI-based prices in case those are in favour of their side. AI-predicted prices might be higher/lower than last month/week as these predictions will depend on external data patterns and historical changes in material prices.

Summary:

I strongly believe we will soon reach to the autonomous negotiations stage. AI based pricing with accurate analytical predictions will help both the manufacturers and their vendors.

Notes:

I want to thank Ceren Ispir and Doruk Ekşi for their strong valuable efforts as Data Science team.

References:

Alalawin, Abdallah & Arabiyat, Laith & Alalaween, Wafa & Qamar, Ahmad & Mukattash, Adnan. (2020). Forecasting vehicle’s spare parts price and demand. Journal of Quality in Maintenance Engineering. ahead-of-print. 10.1108/JQME-03–2020–0019.

--

--

Ozan Çevik
KoçDigital

Data Analytics, AI Business Consultant mainly focusing on Supply Chain