bollinger mark & transform

[mbostock]

Fixes #548.

[mbostock]

Not sure n and k should be required positional arguments here. Probably would be more consistent to have a bollingerOptions argument, and allow a single combined options, too. And maybe a single number could be used to specify n; n defaults to 20 and k defaults to 2.

[mbostock]

By negating k here, bollingerX behaves like mapX, and can be used for either area or line!

[mbostock]

Nice, we get a build failure now when we don’t document something. 😁 I’ll add that to the PR.

[Fil]

Nice! It’s a useful and sound method for general time series analysis.

I'm just wondering if we should keep that name and defaults of the transform. “Bollinger” seems to be confined to finance (according to wikipedia), and has the default multiplier = 2 (not a bad choice, but mathematically surprising).

Reading https://www.bollingerbands.com/bollinger-bands, I see that "bands" is plural—they represent the three lines y = moving average MA, y2 = MA+2 Dev, and y1 = MA-2 Dev. Whereas how I see it in our usage, we're considering a singular band, the area between y1 and y2.

Also, there is a trademark on the name.

Even in finance, people seem to be using “volatility” often (shown as a 1-sigma band). The 1-sigma band excludes 31.8% of the distribution on a normal model, so it makes sense to use a 2-sigma band as the default (4.6% outliers), but it's also legit to do 3-sigma (0.3%), etc. Other data science domains seem to be using words like “rolling window variance” or “rolling window deviation”, and plotting them with 1-sigma, 2-sigma, 3-sigma lines or bands, or adapting the sigmas to match 10%, 5%, 1%….

[Fil]

The only blocking remark is that there is a TODO in the documentation.

Non-blocking remarks & questions:

1. To apply in log-space rather than in linear-space (e.g., to avoid negative values if the multiplier × the deviation is high), one can do:

Plot.bollingerY(aapl, { x: "Date", y: (d) => t(d.Close), k }).plot({
  y: { transform: (d) => t.invert(d), domain: [0, 300], nice: true }
})

with t = d3.scaleLog() 👍. (Note that the difference on aapl with the default parameters is totally imperceptible.)

2. Should we add the anchor option? For time series we obviously don't want to use “values from the future“, and if the data is coming in reverse chronological order (e.g. data = d3.reverse(aapl), it seems important to either anchor: "start" or use an explicit sort: "Date". Similarly, it could be useful to anchor “middle” if x is not time, and we want to gauge local variance.

3. In that vein I'd also want to expose the underlying strict option, which would be particularly useful to set to false when the data has gaps (e.g., interval: "day" on the aapl dataset). (If we support that option and default it to true, we should document it as it is the converse of windowY's default.)