Time series interpolation

General Usage
, ,
1

Suppose i have two vectors, one containing timestamps and one containing values:

using Dates
dt = collect(Date(1900,1,1):Day(2):Date(1902,1,1))
y = randn(length(x))

I want to upsample these data to daily resolution by interpolation. However, I couldn’t find a way to do this yet. I tried interpolating with Interpolations.jl, but when using the dates as index it never finishes. When I use a “helper index” that is just a number range, I get a MethodError, because it can’t interpolate the dates I guess.

using Interpolations
LinearInterpolation(dt, y)  # This command never finishes

i = range(1, len(dt))

LinearInterpolation(i, dt)  # MethodError: no method matching *(::Float64, ::Date)

Ialso tried to put these vectors in a TimeArray from TimeSeries.jl and tried using TimeSeriesResampler:

using TimeSeries, TimeSeriesResampler
ts = TimeArray(dt, y)
resample(ts[:A], Day(1))

This correctly constructs a TimeSeriesResampler object. However, in the documentation I only see options like sum or mean to aggregate data to a lower resolution and no options for interpolation.

Can anyone point me to some ideas on how to interpolate time series data (linearly)?

2

One option is to convert the dates to number of days:

itp = LinearInterpolation(Dates.days.(dt), y)
dt1 = Date(1900,1,1):Day(1):Date(1902,1,1)
y1 = itp(Dates.days.(dt1))

Anoter way of getting number of days as floats is:

(dt .- dt[1])/Day(1)

NB:
As @DNF would say, do not use collect, unless really necessary.

1 Like
3

Great and simple solution. I guess it will also work correctly with irregularly sampled timeseries, yay! Two follow up questions:

  1. How to get back to a date from the integer day values? Dates.Date.(Dates.days.(dt1)) doesn’t work.

  2. Is there a similar thing if the timestamps are sub-daily resolution? Like Dates.seconds.(dt1)? Tried this but the method doesnt exist…

EDIT: Found it. For conversion of dates to days since epoch and vice versa you can use eds = Dates.date2epochdays(Date(1900, 1, 1)) and Dates.epochdays2date(eds). This also works with datetime (useful for timeframes samaller than days) with ems = Dates.datetime2epochms(DateTime(1900, 1, 1)) and Dates.epochms2datetime(ems).

4

There is:

Dates.datetime2epochms.(DateTime.(dt))

which takes DateTime and returns the number of milliseconds since the rounding epoch (0000-01-01T00:00:00).

The inverse is: Dates.epochms2datetime()

The use of these functions is probably preferable.