I think that the most effective strategy might be to first use the Arduino as a data acquisition front-end for a more capable machine for the least squares analysis to discover the specifics of the relationships, then make use of that relationship in the production Arduino software.
When that strategy can be adopted, the need to maintain a statistically-significant collection of data can be much reduced.
I did a bit of that myself this past weekend to come up with a formula for the amount of power to be applied to the ignition heater of a small LENR reactor. I wanted a formula that my controller code could adjust as it "learned" about the quirks of each particular reactor so which it might be connected. I'll attach some plots so you can see what I'm talking about and, perhaps, find the technique useful in your own project.