I've showed already a few cases where heatmap can enhance a line chart decoding. Here is another possible case.
Using sliding windows for encoding the strength of the correlation over time of two variables. This approach allows a nuanced interpretation of the relationships between variables.
True, splitting data makes sense for presidential popularity. But for gasoline, not so. For multiple values in the same month, I used average.
A sliding window for correlation estimation is really helpful, I tried 6, 12, 24, 48 months. 12 gave me the best reflection of my visual perception.
CSP supporters need to come up with something consistent to beat a dual chart with the highlighted correlation. The "legend" was for explaining the color encoding, not a design proposal.
@xangregg
If is a too short window, than is not much stat involved, I am not excluding a variable window size and possible lags to be checked. The "best" inplace correlation coefficient (R) idea:
For a large enough sequence length to be relevant for the given set, N, compute R for all N sequences which include the studied point and choose the one with the largest abs R value. For N=Total Count, all points will score the global correlation.
One can use p-value, too, as the chosen criteria.
@xangregg
Literally using brute force (no convergence/regression assumed) may require too much CPU, yet the fact the number of calculations grows linear with the Sum(TotalCount - WindowSize - Lag_i + 1) with Lag_i varying from 0 to a few, gives a bit of a computational sense in case we want to limit it using increments others than 1.
In fact a 1:1 variation "correlation" is the 2nd graph (orange-purple palette), with the smoothing param controlling the data "noise"...
Still fun, though.
The "fun" part of this exercise is to give a quantifiable/statisticsl touch to our conclusions, hardly possible by visual inspection for variable, ordered/sequential linear correlations, or lags.
Intuitively, CSP supporters are trying to show just that (minus the lag part). I do believe that our visual decoding power needs to be saved for less difficult visual exercises than this. Especially for untrained audiences.
Non linear relationships are easier to grasp in scatterplots.
Daniel Zvinca
Xan Gregg