I'm not sure but here is the best solution I can provide: I feel sort of optimization should be used to solve this issue, or definitely better `model` (rather then linear OLS model) but nonlinear... diff_binAB <- with(data, unique(diff_AB)) mse <- numeric(length(diff_binAB)) for(i in 1:length(diff_binAB)){ pwise <- with(data, lm(retA ~ diff_AB*(diff_AB < diff_binAB[i]) + diff_AB*(diff_AB >= diff_binAB[i]))) mse[i] <- summary(pwise)[6] } mse <- as.numeric(mse) mse diff_binAB[which(mse==min(mse))] # -0.07 diff_binAC <- with(data, unique(diff_AC)) mse1 <- numeric(length(diff_binAC)) for(i in 1:length(diff_binAC)){ pwise <- with(data, lm(retA ~ diff_AC*(diff_AC < diff_binAC[i]) + diff_AC*(diff_AB >= diff_binAC[i]))) mse1[i] <- summary(pwise)[6] } mse1 <- as.numeric(mse1) mse1 diff_binAC[which(mse1==min(mse1))] # 0.04 Here the results would suggest that the `return` (rate of change) is `explained` if the difference between `A` and `B` is at `-0.07` (negative difference) and `0.04` with possitive difference between `A` and `C`.