stjordanis/r-tseries-patterns

trend / momentum and other patterns in financial timeseries

Financial Timeseries Patterns library

This package will contain a collection of price pattern detectors (online and offline). I am starting this library by open sourcing one of the labeling algorithms I use.

A few years ago developed an algorithm to label momentum and trend patterns in intra-day or daily price data. In spite of its simplicity, has performed quite well as compared to a number of more complicated statistical approaches. As is not especially proprietary, hence thought to share this more broadly. I will be adding other pattern related algorithms to this library over time.

I use these algorithms for:

• collecting price moves for pattern analysis
• comparing online trend or MR signal versus optimum behavior as identified ex-post by this labeler
• labels for supervised machine learning in learning momentum signals
• studying market microstructure around large moves

AmplitudeBasedLabeler

The labeler behavior is defined by two parameters (which seem intuitive from a trading perspective):

• minimum trend / momentum amplitude of interest
  • this should be some multiple of volatility / noise
• maximum amount of noise allowed in move:
  • defined by maximum period where no new high (low) is achieved), as well as
  • no drawdown in move exceeding the minimum move amplitude

There are other ways to define noise or extension, but these choises resulted in a super-simple model, that works well. In addition an incremental OLS is performed to determine which points best fit the move, discarding outliers around the edges.

Examples

Below are some examples of the same (intra-day) data series, parameterized for more noise, less noise, higher or lower minimum amplitudes.

Labeling (minamp = 20bps, Tinactive = 5mins)

This was performed on 30sec bars (so 5mins = 10 bars)

labeler <- AmplitudeBasedLabeler$new (minamp = 20, Tinactive = 10)
labels <- labeler$label (df)
labeler$plot()

Labeling (minamp = 20bps, Tinactive = 15mins)

This was performed on 30sec bars (so 15mins = 30 bars)

labeler <- AmplitudeBasedLabeler$new (minamp = 20, Tinactive = 30)
labels <- labeler$label (df)
labeler$plot()