4 Steps: Predictive Analytics in Procurement

Identify interdependencies, support forecasts, detect price risks early on and illustrate savings potentials with correlation and regression calculations

Predictive Analytics in Procurement

Together with the scope and importance of the purchase, the need for statistical methods, which help in researching the cause and interdependencies of price trends, increases. For this purpose, third generation spend management systems, such as SpendControl, offer comprehensive statistical/ mathematical methods.

Price trends will be forecast, future risks avoided and potentials used with the aid of an extensive analysis of past procurement data.

Possible correlations between material prices and price indices (e.g. raw materials such as steel or semi-finished products like basic plastics) are therefore revealed, cost-pushers are identified, and leading indicators are established. These alert a company of cost trends in purchasing ahead of time.

It is one of our objectives to answer the following questions:

• When will the purchasing price be likely to change (price forecast)?
• By means of what indices can it be best forecast?
• What measures need to be derived from the measured tendency of a relevant index?

The first step of pricing analysis: Data Cleansing

The first and most important prerequisite for the great value of a statistical analysis of purchasing prices is the accuracy of the underlying data. For quality assurance purposes, comprehensive error or outlier detection of the purchasing data is conducted in advance. The result is the identification of strong prices, purchasing volumes or spend outliers, their cause(s) and a general estimate of the existing data quality.

Outliers such as these can, among other things, be explained through errors in data input, incorrect conversion rates or mixed quantity units. The cause for severe purchasing price deviations could, however, be simply a change in supplier or different order amounts.

Transaction date	Purchase amount	Supplier	Price
02.02.2013	1000KG	Supplier A	3,00€
05.02.2013	55KG	Supplier A	11,21€ 1
11.02.2013	1070KG	Supplier A	2,97€
15.02.2013	980KG	Supplier A	3,07€
17.02.2013	1000KG	Supplier A	3,03€
20.02.2013	1000KG	Supplier A	10,40BRL 2
25.02.2013	950KG	Supplier A	3,07€
01.03.2013	1070KG	Supplier A	2,97€
06.03.2013	980KG	Supplier A	4320100,51€ 3
07.03.2013	1000KG	Supplier A	3,03€
25.02.2013	950KG	Supplier A	3,07€
01.03.2013	1070KG	Supplier A	2,97€
06.03.2013	970KG	Supplier B	5,60€ 4
07.03.2013	1000KG	Supplier B	5,60€ 4
25.02.2013	950KG	Supplier A	3,07€
01.03.2013	1070KG	Supplier A	2,97€
06.03.2013	975KG	Supplier A	3,03€
07.03.2013	1000KG	Supplier A	3,03€

₁ At this point, an abnormality in the material price was observed. Since an unusually small quantity was ordered at this time, this deviation can be justified and is most probably not an error. ₂ An outlier, which can be traced back to a currency exchange during documentation, was identified here. This abnormality can be corrected through a conversion to Euros. ₃ A severe deviation, as is shown here (with an unchanged currency), is evidently an error that needs correction. ₄ An unusually high material price arises here after separating from “traditional” supplier A. This is most probably not an error, although it should be verified whether the change of suppliers was really necessary at that point in time.

The errors that occurred are corrected depending on the outcome, or are deleted for the following analysis.

Second step: The recognition of correlations between the price and its “drivers”

Valid price analysis can only be conducted after guaranteeing that the data quality is sufficient. To achieve this, a professional assessment regarding the possible external price or cost drivers (e.g. commodity prices, labor costs, economic fluctuations) must be submitted first, in which the data’s validity is verified with a mathematical/ statistical correlation analysis. This way, it is possible to see whether the correct indices are actually being used as indicators or “price drivers”, or whether other cost drivers, which are better suited, exist. Additionally, the methods make it possible to determine if the purchasing price fluctuates within the expected patterns.

The diagram shows a high (linear) correlation between the material price and the trading rate and the currency of the supplier and the currency of the purchaser. With the help of the regression line, a rough estimate of the impact of the trading rate on the material price can be calculated.

Third step: Early warning system for “unauthorized” price deviations

In order to determine at which time and in to which degree the purchasing price has distanced itself from a possible indicator (or forecast price development by way of a “should price”), additional time-limited correlations can be revealed with a detailed procedure. By means of continuous updating, this procedure also serves as an early warning system: in case of possible inconsistencies between price and indices developments, the underlying causes can be researched directly and possible countermeasures can be introduced.

The figure illustrates how the mentioned early warning system operates. A correlation degree between a material price and an expected indicator were measured and documented at intervals over a longer period of time.

Fourth step: Extrapolate forecasts and methods

After the price indicators or cost drivers have been identified, their individual quality must be evaluated. It must be clarified with what loading each relevant index affects the purchasing price (e.g. with the help of a multiple regression), and what risks or potentials are involved. A too strong orientation of the price toward a strongly fluctuating, difficult to foresee index should be viewed just as critically as a nonexistent correlation with the actual expected indicators. The knowledge of such instances, however, gives the company the opportunity to react with appropriate countermeasures or to be ready for future market developments solely through index forecasts.

The diagram shows an index forecast with the help of an ARIMA model (auto-regressive-integrated-moving-advantage). Confidence intervals are formed in order to assess in which are the index score will be situated in the future.

Price forecasts of this kind are also described in literature under the term “Predictive Forecasting”. In the purchasing department, the usability of these kinds of progressive, analytical forecast methods (as described) primarily depends on:

• that a sufficient and fully completed number of price time series are available, which
• feature, if possible, very few errors in terms of outliers or negative values (e.g. cancellations or reversals)

A further challenge when measuring cost driver effects are time impact delays. This way, for example, a change in oil pricing influences the purchase price of basic plastics only after a delay of several weeks ( the own warehousing not taken into consideration). To identify this effect, diverse time shifts are applied during the correlation analysis. This means that the possible price indicator is moved along the timeline several times, and after each shift, a check for a correlation with the material price is carried out. This ensures that delayed interdependencies are always recognized as well.

A delay as such can be very helpful for price forecasts. If, for example, a correlation between a material price and a measureable indicator is detected after a shift of one month, an approximate forecast of the material costs in the following month can be made based on this indicator.

Here we are able to see the impact that the change in oil prices (Brent) has on the polymer index. The first graph still depicts a relatively strong dispersion (weak correlation). If, however, we take a time impact delay of one month into consideration (second graph), a rough linear correlation becomes evident.