A Scala / Python library for interacting with time series data on Apache Spark.

Docs are available at http://cloudera.github.io/spark-timeseries. Scaladoc is available at http://cloudera.github.io/spark-timeseries/scaladocs/index.html.

The aim here is to provide

• A set of abstractions for manipulating large time series data sets, similar to what's provided for smaller data sets in Pandas, Matlab, and R's zoo and xts packages.
• Models, tests, and functions that enable dealing with time series from a statistical perspective, similar to what's provided in StatsModels and a variety of Matlab and R packages.

The library is geared towards use cases in finance (munging tick data, building risk models), but intends to be general enough that other fields with continuous time series data, like meteorology, can make use of it.

The library currently expects that individual univariate time series can easily fit in memory on each machine, but that collections of univariate time series may need to be distributed across many machines. While time series that violate this expectation pose a bunch of fun distributed programming problems, they don't tend to come up very often in finance, where an array holding a value for every minute of every trading day for ten years needs less than a couple million elements.

The library sits on a few other excellent Java and Scala libraries.

• Breeze for NumPy-like, BLAS-able linear algebra.
• JodaTime for dates and times.
• Apache Commons Math for general math and statistics functionality.
• Apache Spark for distributed computation with in-memory capabilities.

The central abstraction of the library is the TimeSeriesRDD, a lazy distributed collection of univariate series with a conformed time dimension. It is lazy in the sense that it is an RDD: it encapsulates all the information needed to generate its elements, but doesn't materialize them upon instantiation. It is distributed in the sense that different univariate series within the collection can be stored and processed on different nodes. Within each univariate series, observations are not distributed. The time dimension is conformed in the sense that a single DateTimeIndex applies to all the univariate series. Each univariate series within the RDD has a key to identify it.

TimeSeriesRDDs then support efficient series-wise operations like slicing, imputing missing values based on surrounding elements, and training time-series models. In Scala:

val tsRdd: TimeSeriesRDD = ...

// Find a sub-slice between two dates 
val subslice = tsRdd.slice(new DateTime("2015-4-10"), new DateTime("2015-4-14"))

// Fill in missing values based on linear interpolation
val filled = subslice.fill("linear")

// Use an AR(1) model to remove serial correlations
val residuals = filled.mapSeries(series => ar(series, 1).removeTimeDependentEffects(series))

In Python:

tsrdd = ...

# Find a sub-slice between two dates
subslice = tsrdd['2015-04-10':'2015-04-14']

# Fill in missing values based on linear interpolation
filled = subslice.fill('linear')

• Aligning
• Slicing by date-time
• Missing value imputation

• Exponentially weighted moving average (EWMA) models
• Autoregressive integrated moving average (ARIMA) models
• Generalized autoregressive conditional heteroskedastic (GARCH) models
• Missing data imputation
• Augmented Dickey-Fuller test
• Durbin-Watson test
• Breusch-Godfrey test
• Breusch-Pagan test

• Multivariate T Distribution

Value at Risk (VaR) and Expected Shortfall (CVaR) through

• Monte Carlo simulation
• Bootstrapped historical simulation

We use Maven for building Java / Scala. To compile, run tests, and build jars:

mvn package

To run a spark-shell with spark-timeseries and its dependencies on the classpath:

spark-shell --jars target/sparktimeseries-0.0.1-jar-with-dependencies.jar

To run Python tests (requires nose):

cd python
nosetests

To publish docs, easiest is to clone a separate version of this repo in some location we'll refer to as DOCS_REPO. Then:

mvn site -Ddependency.locations.enabled=false
mvn scala:doc
cp -r target/site/* $DOCS_REPO
cd $DOCS_REPO
git checkout gh-pages
git add -A
git commit -m "Some message that includes the hash of the relevant commit in master"
git push origin gh-pages