So far the only feedback that I have received on the tidalGauge/ENSO correlation concerns applying a 3 month offset-delay to the tidal data data before doing the correlation. Two others elsewhere mentioned this, and Nick's suggestion was to use Sydney atmospheric pressure data instead, which I inferred meant that a time shift $\delta T$ would be unnecessary. I am still looking for the Sydney pressure data, but in the interim I decided to show how subtle this 3-month effect is.

This was what was fed into the optimizer:

$$ SOI(t) = k (Tide(t - \delta T) - Tide(t - \delta T - \Delta T)) $$

In the figure below, the top chart has the delay and the bottom has no offset delay, $\delta T =0$, with $\Delta T$ unchanged at 23 months. The difference in correlation coefficient is 0.65 vs 0.62. The correlation coefficient was used as the measure to minimize against via the optimizer and the 3-month offset and the 23-month delay difference is what the machine learning determined as the best correlation.

![correlation](http://imageshack.com/a/img631/488/J8RcwQ.gif)

I don't think this offset delay is the most interesting feature in the correlation, but since three people made essentially the same observation I figured I would dive a little deeper into the optimization approach. Bottom-line is that I am not putting a finger on the scale, and this is what the machine learning discovered.

This was what was fed into the optimizer:

$$ SOI(t) = k (Tide(t - \delta T) - Tide(t - \delta T - \Delta T)) $$

In the figure below, the top chart has the delay and the bottom has no offset delay, $\delta T =0$, with $\Delta T$ unchanged at 23 months. The difference in correlation coefficient is 0.65 vs 0.62. The correlation coefficient was used as the measure to minimize against via the optimizer and the 3-month offset and the 23-month delay difference is what the machine learning determined as the best correlation.

![correlation](http://imageshack.com/a/img631/488/J8RcwQ.gif)

I don't think this offset delay is the most interesting feature in the correlation, but since three people made essentially the same observation I figured I would dive a little deeper into the optimization approach. Bottom-line is that I am not putting a finger on the scale, and this is what the machine learning discovered.