Problem Statement & Challenges:
The client is a multinational automobile manufacturer, based in Europe. They are one of the biggest producers of premium cars.
The client’s procurement and supplier quality team forecast the price of commodity and use these forecasts for budget planning and vendor negotiations for the subsequent years. The team used qualitative or judgmental techniques to forecast the price of commodity . They made educated guesses based on intuition, knowledge and experience.
High volatility in the price of commodities due to different factors like the exchange market, political and geographical events which lead to a lot of planning inefficiency and budget gaps. Therefore, the client wanted to seek a data-driven approach to forecast the commodity prices.
A major challenge was with the battery-grade commodities like Lithium and Cobalt as their prices have been extremely volatile. The other challenge was the unavailability of import/export data, global inventory levels, and other commodity-related prices.
The figure above indicated the different challenges in the project
Commodity price forecasting is challenging as it requires an in-depth domain knowledge of commodities and a thorough understanding of both qualitative and quantitative approach of forecasting.
To achieve it, this approach was followed:
The figure indicates the step-by-step approach in solving the problem
- Historical data study helped in identifying price trends, benchmarks and impact of various global events on the price of the commodity.
- Commodity prices are affected by external events, which can be captured by identifying peaks or change in trend/pattern in price data.
- Key drivers of commodity price can be identified by understanding the evolution of commodity into the market.
- Impact of events and drivers on commodity price can be understood by using statistical techniques likes correlation, hypothesis testing and generalized linear model (GLM) and the impact, in turn, can be used as a reference when such events occur in future.
- Quantitative forecasting is broadly classified into univariate, multivariate and advanced techniques. Due to the limited amount of data and monthly level granularity, it is ideal to use univariate or multivariate forecasting techniques.
- Although there are many methods used for accuracy calculation, MAPE (Mean Absolute Percentage Error) is used in time series forecasting because it is independent of the scale of the output variable. MAPE for Naïve method and other methods was calculated across different time stamps and finally, the model which showed better performance across time was selected.
Below is the list of univariate techniques used:
- ARIMA (Autoregressive Integrated Moving Average)
- SARIMA (Seasonal Autoregressive Integrated Moving Average)
- ARCH/GARCH (Autoregressive Conditional Heteroskedasticity)
- Simple Exponential Smoothing
- Holt-Winters Double Exponential Smoothing
- Holt-Winters Triple Exponential Smoothing
- Univariate Unobserved Component Model
- SETAR (Self-Exciting Threshold Autoregressive)
- STAR (Smooth Transition Autoregressive)
Below is the list of multivariate techniques used:
- ARIMAX (Autoregressive Integrated Moving Average with Exogenous Variable)
- SARIMAX (Seasonal Autoregressive Integrated Moving Average with Exogenous Variable)
- M-UCM (Multivariate Unobserved Component Model)
- VARIMA (Vector Autoregressive Integrated Moving Average)
- VECM (Vector Error Correction Model)
- A Tableau dashboard was created so that the procurement and supplier quality team can easily consume the output. The dashboard had multiple forecasts, comparison of the forecasts with the expert judgemental forecasts and insights on the past price trends.
- While the forecasts were relatively better than the expert forecasts, procurement managers still had to rely on observing the market behaviour to look out for any geopolitical events, economic events or any other news that would directly impact the price of the commodity.
To account for this, a new feature was developed in the dashboard that actively lists the news from various websites and displays the headline, which on clicking, would redirect to the news article. This feature would help the procurement manager to understand whether the price of the commodity will increase or decrease and based on the judgement he can modify the forecast.
After applying the different univariate and multivariate techniques the best model was selected and used for forecasting.
- The forecasting framework helped improve accuracy by 10 – 30 percent basis points across commodities, with ability to explain fluctuations with external events.
- The procurement teams are now able to plan their budgets more accurately and negotiate better prices with vendors, leading to significant cost savings.
Working with the team was a great experience. The team was very flexible even when the scope of the project expanded. The team worked hard to get such fantastic results.”Manager, Processes & Systems Procurement