Spurious regression between two futures with the same underlying highly correlated (cor=0.9)

analyzing the correlation between soybean and soybean meal futures in ECBOT, and making a linear regression in R between them I check with an ADF Test that the residuals are not stationary, so theoretically both series are not cointegrated, else they have a spurious relationship. But, has this sense? They are over the same underlying and highly correlated (not like Soybean Oil and Soybean futures that have poor correlation).

Cheers, Gonzalo.

• Just thinking out loud: If (1) soybean = soybean meal + soybean oil and (2) soybean and soybean oil are not cointegrated, then soybean cannot be cointegrated with soybean meal as that would mean the two sides of the equation above would not cointegrate, and thus the equality could not hold. But perhaps (1) does not hold (even approximately)? Also, could you include a graph with the two series? Perhaps we could see something in it that justifies the outcome. – Richard Hardy Mar 12 '17 at 12:36
• <a href="ibb.co/kA4pAa"><img src="image.ibb.co/eszd3v/…" alt="Captura_de_pantalla_2017_03_13_a_las_22_53_47" border="0"></a> <a href="ibb.co/f8Ny3v"><img src="image.ibb.co/jO0kiv/…" alt="Captura_de_pantalla_2017_03_13_a_las_22_54_14" border="0"></a> – Gonzalo Federico Mar 14 '17 at 2:12
• Ignore all before, I hate this editor. I dont have enough points for writing a complete answer to your comment so I will say that in the first image there are the both series multiplying ZM with the lineal coefficient of the regression (1.915) having a correlation of 0.9. ibb.co/kA4pAa After reading your comment I found the next serie ZS = ZM + ZL ibb.co/h4aROv Now, there is no stationary again, with non stationary residuals in the regression, but the graph appears to be more predictable between ranges and the correlation is more strong (0.93 aprox) Cheers – Gonzalo Federico Mar 14 '17 at 2:33
• The first analysis was done in a frequency of days in year, but after checking this multiple regression in other frequencies (days in month, minutes in day, and minutes in week) the correlation get worst and there is no stable range in the spread (and obviously there is no stationarity). Maybe trying with another non-linear regression method or with a non parametric method could work... any ideas? – Gonzalo Federico Mar 14 '17 at 11:01