Regression Analysis

The correlation analysis of section 4.1.1 showed some degree of interdependence of planned maintenance activities and product price spreads. However, to properly describe a dependency in order to confirm Hypothesis 1A, a regression analysis with explanatory variables is needed (Read, 1998).

To measure the contribution of maintenance activity to the overall explanatory power, crude oil price and GDP growth are included in the regression model. The latter ones are common business cycle indicators (Abberger & Nierhaus, 2011) and crude oil is, in addition, part of the plastic value chain.

Three results are expected which are not reflected in the correlation analysis but can be visible in a regression analysis which unveils explanatory dependency: First, spot markets could in theory be a good indicator of market dynamics caused by large-scale maintenance-activities.

However, liquidity on ethylene and propylene spot markets is low encountering days without a

single trade (see 1.4). It is unlikely that market dynamics such as maintenance activity is properly reflected in spot market prices due to the low volumes traded. Hence, lower explanatory power is expected for the model predicting the spot prices than the ones predicting contract price. Second, the retrieved dataset gives the opportunity to observe a natural experiment¹¹. The environment for the upstream plastic value chain was changed due to the impact of the Lehman Shockwave in two ways: a) Lower level of vertical integration and b) monthly instead of quarterly settled contract prices for ethylene and polyethylene. In response to the Lehman Shock, several big industry players sold their polymer units (see 1.1.1). Although long-term purchase contracts exist between the former integrated parties, S&OP alignment and communication might be affected negatively. As price volatility increased (see 1.4.2) this would mean that although the economic pressure for the petrochemical industry in Europe to rationalise has increased since the Lehman Shock, the supply chains have become less efficient and responsive. This is under the assumption that former integrated entities perform worse in planning synchronisation when they are separated. In addition, in the earlier period from 2005 to 2008, ethylene contract prices were settled on a quarterly basis in contrast to monthly settled prices in the second period (see 1.4). This less frequent settling could dilute maintenance impact statistically. Consequently, the predictive power of the maintenance variable is expected to be higher in the second period (2009 to 2012) than in the first period (2005 to 2008) of the natural experiment. Third, as butadiene price spread correlations are significant and in comparable magnitude to those of ethylene and propylene, predictive power of regression models is expected to be akin to those for ethylene and propylene.

For all clearly significant correlation relations of 4.1.1 a multiple regression analysis was carried out. The dependent variable price spread is the selected price spread between ethylene, propylene as well as butadiene and naphtha.

To assess the full predictive power of maintenance activities on relevant price spreads, three alternative models are compared. First, only planned (the total planned maintenance activity in production loss in kilotons of ethylene) is added to the null-model. Second, a model commonly used in the industry with the two independent variables crude oil (the price for one barrel crude oil Brent FOB North Sea in €) and GDP (the quarterly reported change in GDP reported in full decimals¹²) is investigated. Third, a combined model with all three predictors is studied.

𝑀𝑜𝑑𝑒𝑙 1:

𝑝𝑟𝑖𝑐𝑒 𝑠𝑝𝑟𝑒𝑎𝑑₁ = 𝑏₀+ 𝑏₁∙ 𝑝𝑙𝑎𝑛𝑛𝑒𝑑 _{EQ. 4.1}

𝑀𝑜𝑑𝑒𝑙 2:

𝑝𝑟𝑖𝑐𝑒 𝑠𝑝𝑟𝑒𝑎𝑑₁ = 𝑏₀+ 𝑏₁∙ 𝑐𝑟𝑢𝑑𝑒 𝑜𝑖𝑙+𝑏₂∙ 𝐺𝐷𝑃 _{EQ. 4.2} 𝑀𝑜𝑑𝑒𝑙 3:

𝑝𝑟𝑖𝑐𝑒 𝑠𝑝𝑟𝑒𝑎𝑑1 = 𝑏0+ 𝑏1∙ 𝑝𝑙𝑎𝑛𝑛𝑒𝑑 + 𝑏2∙ 𝑐𝑟𝑢𝑑𝑒 𝑜𝑖𝑙 + 𝑏3∙ 𝐺𝐷𝑃 _{EQ. 4.3} Equation Eq. 4.1, Eq. 4.2 and Eq. 4.3 give the structural model for the eight different price spreads. Table 3 compares the models for the different datasets (time periods). For each price spread, the R² and adjusted R² values of the three models are given. In addition, the significance of the F-ratio is reported.

11 “A natural experiment is an empirical study in which the experimental conditions (i.e., which units receive which treatment) are determined by nature or by other factors out of the control of the experimenters and yet the treatment assignment process is arguably exogenous.” (The New Palgrave Dictionary of Economics, 2013).

12 A GDP variable value of 1 represents a GDP increase of 1%.

25 Maintenance-Price Model Two model sets are highlighted in Table 3: The set explaining the ethylene/naphtha contract price spread and the one explaining the propylene/naphtha contract price spreads. The two products are the most important outputs of a cracker and the contract prices determine a large amount of the overall margin achieved in cracker operations. The given results show that planned maintenance activities alone can explain 24.3% (respectively 17.3%) of the variance in the contract price spreads and hence confirm the presumed significant impact of maintenance activities on the market. Adding the previous month’s price spread as a control variable to the other predictors increases the R²-value but does not lower the predictive power of planned. The standardized Beta-values¹³ are in the same magnitude as those for planned and crude oil (see Table C-4). This is additional evidence that including scheduled maintenance in a price spread forecast adds significant value. Consequently, Hypothesis 1A: Maintenance activities of cracker units have a positive effect on price spreads of commodity products, can be accepted. Figure 18 illustrates the added value of including planned maintenance activities in prediction models. No other models show strong significant effects of maintenance activities comparable to those for ethylene and propylene. As expected, the models for ethylene and propylene contract prices solely based on maintenance activity are only significant since those prices are settled monthly (2009 to 2012).

FIGURE 18: ADDED VALUE OF PLANNED VARIABLE

An additional test of Model 3 for the ethylene spread with reversed order yielded the exact same R²-values: First, crude oil was added to the model and then planned. Table C-5 shows the model summary. This result is due to the fact both variables are not correlated (see Table 2).

13 planned: .311, crude oil: .352, GDP: .046 (non sig.), spread-1: .330 + 28%

+ 17%

.000 .200 .400 .600 .800

Ethylene spread [contract]

Propylene spread [contract]

R²-values (2009-12)

crude oil + GDP

added value of planned maintenance

DEPENDENT VARIABLE

bold: significant F-Change at 0.05 level bold italic: significant F-Change at 0.01 level

TABLE 3: STEPWISE REGRESSION ANALYSIS ON DEPENDENT PRICE SPREADS