Yesterday, I presented with Lisa Reisman on a MetalMiner/Spend Matters webinar looking at new sourcing strategies for organizations that purchase material quantities of commodities or raw materials (or in cases where their suppliers do it on their behalf). The central theme of the webinar was how to leverage price index information at the core of five new sourcing strategies. During the discussion, we talked about how underlying third-party datasets (including price indexes) can serve as a foundation for advanced sourcing strategies. The sourcing approaches that leverage underlying price index information focus on leveraging data to:
- Form the basis to run statistical regression models, predictive models
- Identify corollary materials and to better understand the interplay amongst
raw materials and impact on metal price (e.g. FeCr on Stainless, Iron ore and scrap on steel etc)
- Use as a contracting tool (price escalator/de-escalator clauses)
- Embed with existing business intelligence, ERP and sourcing/procurement systems to conduct more sophisticated and robust spend analytics (e.g., spend analysis, sourcing optimization and commodity management platforms)
- To use as a negotiating tactic with suppliers particularly when creating contracts in which metal prices float but value-add premiums are bid competitively and held fixed
Today we'll tackle the first item on the list: the role of data in running predictive models for procurement. During this part of the talk, Lisa began her discussion by sharing a range of variables and forecasting techniques, courtesy of material from MetalMiner partner the Smart Cube (who presented at our Commodity EDGE event earlier in the year). These forecasting techniques and variables include normality, non-stationarity (trends, seasonability), regression models, time series modeling and multicollinearity (a situation when two or more predictor variables in a regression model have tight correlation).
Next, moving onto MetalMiner and MetalMiner IndX-specific material, Lisa gave an example of a particular regression we ran internally looking at a predictive model for steel pricing that achieve an r-squared value of over 0.96. The purpose of this model was to highlight all the potential elements that can go into a successful predictive model. In this case, we included such variables as steel scrap prices, cement and concrete productions production, housing starts, employment, automobile assembly production and capacity/utilization within the sector. In our model taking in multiple years of data, we found that certain factors had a tighter correlation than others with predicting future prices (MetalMiner indX information on past pricing was an essential component of the model).
We then shared a specific commodity forecasting dashboard that we built for copper to highlight the importance of considering a range of factors in any type of forecasting model -- from basic regression analysis to more advanced techniques. For copper, it is important to consider a broad set of inputs in a predictive model ranging from past index pricing data on a local/regional level in key markets, demand, currency valuations (dollar/euro/RMB), supply factors/capacity utilization, speculative activity (volume in/off warrants), China GDP, CPI and PMI data, electricity costs, global auto sales, ETF inflows and other related variables.
The concept of an interactive forecasting dashboard, even starting with charting in Excel as a foundation, is to enable an interface that encourages drilling into multiple sets of data in context to understand not only what correlations appear to work, but when correlations break down (and under what circumstances). A model may have a very high degree of accuracy, but knowing the indicators that can suggest when it may break (e.g., China PMI dropping below a certain level) is just as important as getting the predictive capability right in the first place under stable market circumstances. Which of course starts with leveraging the right sets of market index data, including commodity-pricing indexes to feed the forecasting model.
Next week, we'll continue to explore this topic in more detail, looking at how to use price index data in spend analysis and e-sourcing/optimization environments to drive more effective sourcing strategies.