tshistory 0.15.0

Timeseries store with version control

TSHISTORY

This is a library to store/retrieve pandas timeseries to/from a postgres database, tracking their successive versions.

Introduction

Purpose

tshistory is targetted at applications using time series where backtesting and cross-validation are an essential feature.

It provides exhaustivity and efficiency of the storage, with a simple Python api.

It can be used as a building block for machine learning, model optimization and validation, both for inputs and outputs.

Principles

There are many ways to represent timeseries in a relational database, and tshistory provides two things:

• a base python API which abstracts away the underlying storage

• a postgres model, which emphasizes the compact storage of successive states of series

The core idea of tshistory is to handle successive versions of timeseries as they grow in time, allowing to get older states of any series.

Basic usage

Starting with a fresh database

You need a postgresql database. You can create one like this:

 createdb mydb

Then, initialize the tshistory tables, like this:

 tsh init-db postgresql://me:password@localhost/mydb

From this you're ready to go !

Creating a series

However here's a simple example:

 >>> import pandas as pd
 >>> from tshistory.api import timeseries
 >>>
 >>> tsa = timeseries('postgres://me:password@localhost/mydb')
 >>>
 >>> series = pd.Series([1, 2, 3],
 ...                    pd.date_range(start=pd.Timestamp(2017, 1, 1),
 ...                                  freq='D', periods=3))
 # db insertion
 >>> tsa.update('my_series', series, '[e-mail]')
 ...
 2017-01-01    1.0
 2017-01-02    2.0
 2017-01-03    3.0
 Freq: D, Name: my_series, dtype: float64

 # note how our integers got turned into floats
 # (there are no provisions to handle integer series as of today)

 # retrieval
 >>> tsa.get('my_series')
 ...
 2017-01-01    1.0
 2017-01-02    2.0
 2017-01-03    3.0
 Name: my_series, dtype: float64

Note that we generally adopt the convention to name the time series api object tsa.

Updating a series

This is good. Now, let's insert more:

 >>> series = pd.Series([2, 7, 8, 9],
 ...                    pd.date_range(start=pd.Timestamp(2017, 1, 2),
 ...                                  freq='D', periods=4))
 # db insertion
 >>> tsa.update('my_series', series, '[e-mail]')
 ...
 2017-01-03    7.0
 2017-01-04    8.0
 2017-01-05    9.0
 Name: my_series, dtype: float64

 # you get back the *new information* you put inside
 # and this is why the `2` doesn't appear (it was already put
 # there in the first step)

 # db retrieval
 >>> tsa.get('my_series')
 ...
2017-01-01    1.0
2017-01-02    2.0
2017-01-03    7.0
2017-01-04    8.0
2017-01-05    9.0
Name: my_series, dtype: float64

It is important to note that the third value was replaced, and the two last values were just appended. As noted the point at 2017-1-2 wasn't a new information so it was just ignored.

Retrieving history

We can access the whole history (or parts of it) in one call:

 >>> history = tsa.history('my_series')
 ...
 >>>
 >>> for idate, series in history.items(): # it's a dict
 ...     print('insertion date:', idate)
 ...     print(series)
 ...
 insertion date: 2018-09-26 17:10:36.988920+02:00
 2017-01-01    1.0
 2017-01-02    2.0
 2017-01-03    3.0
 Name: my_series, dtype: float64
 insertion date: 2018-09-26 17:12:54.508252+02:00
 2017-01-01    1.0
 2017-01-02    2.0
 2017-01-03    7.0
 2017-01-04    8.0
 2017-01-05    9.0
 Name: my_series, dtype: float64

Note how this shows the full serie state for each insertion date. Also the insertion date is timzeone aware.

Specific versions of a series can be retrieved individually using the get method as follows:

 >>> tsa.get('my_series', revision_date=pd.Timestamp('2018-09-26 17:11+02:00'))
 ...
 2017-01-01    1.0
 2017-01-02    2.0
 2017-01-03    3.0
 Name: my_series, dtype: float64
 >>>
 >>> tsa.get('my_series', revision_date=pd.Timestamp('2018-09-26 17:14+02:00'))
 ...
 2017-01-01    1.0
 2017-01-02    2.0
 2017-01-03    7.0
 2017-01-04    8.0
 2017-01-05    9.0
 Name: my_series, dtype: float64

It is possible to retrieve only the differences between successive insertions:

 >>> diffs = tsa.history('my_series', diffmode=True)
 ...
 >>> for idate, series in diffs.items():
 ...   print('insertion date:', idate)
 ...   print(series)
 ...
 insertion date: 2018-09-26 17:10:36.988920+02:00
 2017-01-01    1.0
 2017-01-02    2.0
 2017-01-03    3.0
 Name: my_series, dtype: float64
 insertion date: 2018-09-26 17:12:54.508252+02:00
 2017-01-03    7.0
 2017-01-04    8.0
 2017-01-05    9.0
 Name: my_series, dtype: float64

You can see a series metadata:

 >>> tsa.update_metadata('series', {'foo': 42})
 >>> tsa.metadata('series')
 {foo: 42}

Staircase series

A staircase series can be defined as a series of which values originate from successive revisions with a fixed time span between revision date and value date. This is especially useful for backtesting.

Basic staircase

Let us take an example assuming a series called daily_series has been created with insertions given by the following table (considering row indices are value dates, and columns indices are insertion dates):

	2020-01-01 00:00+00	2020-01-02 00:00+00	2020-01-03 00:00+00
2020-01-01	1.1
2020-01-02	2.1	2.2
2020-01-03	3.1	3.2	3.3
2020-01-04		4.2	4.3
2020-01-05			5.3

Supposing this series is a forecast published on a daily basis, we can for example reconstruct the day-ahead forecast series, i.e. the values such that the time span between revision date and value date is 1 day (or more) as follows:

 >>> tsa.staircase('daily_series',
                   from_value_date=pd.Timestamp('2020-01-01'),
                   to_value_date=pd.Timestamp('2020-01-07'),
                   delta=pd.Timedelta(days=1))
 ...
 2020-01-02    2.1
 2020-01-03    3.2
 2020-01-04    4.3
 2020-01-05    5.3
 Name: daily_series, dtype: float64

The name "staircase" refers to the way in which these values are picked from the history:

	2020-01-01 00:00+00	2020-01-02 00:00+00	2020-01-03 00:00+00
2020-01-01
2020-01-02	2.1
2020-01-03		3.2
2020-01-04			4.3
2020-01-05			5.3

Now if instead we consider an hourly forecast series, we may want to define day-ahead forecast as a staircase series with a daily revision occurring at 9am, and link each revision to the 24 hours of the next day. More generally we may want to reconstruct a staircase series where successive revisions each relate to several value dates. Such cases should instead be handled using the block_staircase method described below.

Block staircase

Let us take another example considering the series hourly_series with following insertions:

	2020-01-01 06:00+00	2020-01-01 14:00+00	2020-01-02 06:00+00	2020-01-02 14:00+00
2020-01-01 00:00+00	1.1	1.2
2020-01-01 08:00+00	2.1	2.2
2020-01-01 16:00+00	3.1	3.2
2020-01-02 00:00+00	4.1	4.2	4.3	4.4
2020-01-02 08:00+00	5.1	5.2	5.3	5.4
2020-01-02 16:00+00	6.1	6.2	6.3	6.4
2020-01-03 00:00+00	7.1	7.2	7.3	7.4
2020-01-03 08:00+00	8.1	8.2	8.3	8.4
2020-01-03 16:00+00	9.1	9.2	9.3	9.4
2020-01-04 00:00+00			10.3	10.4
2020-01-04 08:00+00			11.3	11.4
2020-01-04 16:00+00			12.3	12.4

Then the day-ahead forecast with revisions at 9am can be computed as follows:

 >>> tsa.block_staircase('hourly_series',
                         from_value_date=pd.Timestamp('2020-01-01', tz="utc"),
                         to_value_date=pd.Timestamp('2020-01-05', tz="utc"),
                         revision_freq={'days': 1},
                         revision_time={'hour': 9},
                         revision_tz='utc',
                         maturity_offset={'days': 1},
                         maturity_time={'hour': 0})
 ...
 2020-01-02 00:00:00+00:00   4.1
 2020-01-02 08:00:00+00:00   5.1
 2020-01-02 16:00:00+00:00   6.1
 2020-01-03 00:00:00+00:00   7.3 
 2020-01-03 08:00:00+00:00   8.3 
 2020-01-03 16:00:00+00:00   9.3 
 2020-01-04 00:00:00+00:00   10.4
 2020-01-04 08:00:00+00:00   11.4
 2020-01-04 16:00:00+00:00   12.4
 Name: hourly_series, dtype: float64

Note that with revision_time={'hour': 9}, the method ends up picking values from the two 6am insertions (except for the values of 2020-01-04 when latest available revision is 2020-01-02 14:00). Taking revision time after 2pm, say revision_time={'hour': 20}, would instead select values from the 2pm insertions only.

In general, the arguments of block_staircase should be used as follows:

• from_value_date and to_value_date: time range on which values are retrieved
• revision_freq: revision frequency, as a dictionary of integers of which keys must be taken from ["years", "months", "weeks", "bdays", "days", "hours", "minutes", "seconds"]
• revision_time: revision time, as a dictionary of integers of which keys should be taken from ["year", "month", "day", "weekday", "hour", "minute", "second"]. It is used for revision date initialisation. The next revision dates are then obtained by successively adding revision_freq.
• revision_tz: time zone in which revision date and time are expressed
• maturity_offset: time span between each revision date and start time of related block of values, as dictionary of integers. Its keys must be taken from ["years", "months", "weeks", "bdays", "days", "hours", "minutes", "seconds"]. No lag is considered if it is not specified, i.e. the revision date is the block start date
• maturity_time: start time of each block, as a dictionary of integers of which keys should be taken from ["year", "month", "day", "hour", "minute", "second"]. The start date of each block is thus obtained by adding maturity_offset to revision date and then applying maturity_time. If not specified block start date is just the revision date shifted by maturity_offset

Other use cases

The block_staircase method covers multiple use cases, such as week-ahead revisions or revision by business day, as described in the following examples.

Week-ahead staircase

Consider a series named weekly_series with following insertions:

	2021-01-05 (Tue)	2021-01-07 (Thu)	2021-01-12 (Tue)	2021-01-14 (Thu)
2021-01-11 (Mon)	1.1	1.2
2021-01-12 (Tue)	2.1	2.2
2021-01-13 (Wed)	3.1	3.2
2021-01-14 (Thu)	4.1	4.2
2021-01-15 (Fri)	5.1	5.2
2021-01-16 (Sat)	6.1	6.2
2021-01-17 (Sun)	7.1	7.2
2021-01-18 (Mon)	8.1	8.2	8.3	8.4
2021-01-19 (Tue)	9.1	9.2	9.3	9.4
2021-01-20 (Wed)	10.1	10.2	10.3	10.4
2021-01-21 (Thu)			11.3	11.4
2021-01-22 (Fri)			12.3	12.4
2021-01-23 (Sat)			13.3	13.4
2021-01-24 (Sun)			14.3	14.4
2021-01-25 (Mon)			15.3	15.4
2021-01-26 (Tue)			16.3	16.4
2021-01-27 (Wed)			17.3	17.4

Then the week-ahead staircase with weekly revision on Friday can be retrieved as follows:

 >>> tsa.block_staircase('weekly_series',
                         from_value_date=pd.Timestamp('2021-01-10'),
                         to_value_date=pd.Timestamp('2021-01-30'),
                         revision_freq={'days': 7},
                         revision_time={'weekday': 4},
                         revision_tz='utc',
                         maturity_offset={'days': 3},
                         maturity_time={'hour': 0})
 ...
 2021-01-11   1.2
 2021-01-12   2.2
 2021-01-13   3.2
 2021-01-14   4.2
 2021-01-15   5.2
 2021-01-16   6.2
 2021-01-17   7.2
 2021-01-18   8.4
 2021-01-19   9.4
 2021-01-20   10.4
 2021-01-21   11.4
 2021-01-22   12.4
 2021-01-23   13.4
 2021-01-24   14.4
 2021-01-25   15.4
 2021-01-26   16.4
 2021-01-27   17.4
 Name: weekly_series, dtype: float64

It is also possible to retrieve a month-ahead staircase series taking instead revision_freq={'months': 1} and, for example, revision_time={'day': 15} to perform monthly revision every 15th day of the month.

Revision by business day

The block_staircase method allows to express revision frequency and/or maturity time span in business days. Consider a series named business_day_series with these insertions:

	2021-01-13 (Wed)	2021-01-14 (Thu)	2021-01-15 (Fri)	2021-01-16 (Sat)	2021-01-17 (Sun)	2021-01-18 (Mon)
2021-01-13 (Wed)	3.1
2021-01-14 (Thu)	4.1	4.2
2021-01-15 (Fri)	5.1	5.2	5.3
2021-01-16 (Sat)	6.1	6.2	6.3	6.4
2021-01-17 (Sun)		7.2	7.3	7.4	7.5
2021-01-18 (Mon)			8.3	8.4	8.5	9.6
2021-01-19 (Tue)				9.4	9.5	11.6
2021-01-20 (Wed)					10.5	12.6
2021-01-21 (Thu)						13.6

Then we can retrieve a business-day-ahead staircase series with revision every business day as follows:

 >>> tsa.block_staircase('business_day_series',
                         from_value_date=pd.Timestamp('2021-01-13'),
                         to_value_date=pd.Timestamp('2021-01-21'),
                         revision_freq={'bdays': 1},
                         revision_tz='utc',
                         maturity_offset={'bdays': 1})
 ...
 2021-01-14   4.1
 2021-01-15   5.2
 2021-01-16   6.2
 2021-01-17   7.2
 2021-01-18   8.3
 2021-01-19   11.6
 2021-01-20   12.6
 2021-01-21   13.6
 Name: weekly_series, dtype: float64

Optimizing staircase computations with history cache

It may be useful in some cases to compute multiple staircase series from the same source series. For example, given a series named "my_forecast", we could reconstruct both the one-day-ahead and two-days-ahead staircase series by doing

 >>> ts_1da = tsa.block_staircase('my_forecast',
                                  from_value_date=pd.Timestamp('2020-01-01', tz="utc"),
                                  to_value_date=pd.Timestamp('2022-01-01', tz="utc"),
                                  revision_freq={'days': 1},
                                  revision_time={'hour': 9},
                                  revision_tz='utc',
                                  maturity_offset={'days': 1},
                                  maturity_time={'hour': 0})
 >>> ts_2da = tsa.block_staircase('my_forecast',
                                  from_value_date=pd.Timestamp('2020-01-01', tz="utc"),
                                  to_value_date=pd.Timestamp('2022-01-01', tz="utc"),
                                  revision_freq={'days': 1},
                                  revision_time={'hour': 9},
                                  revision_tz='utc',
                                  maturity_offset={'days': 2},
                                  maturity_time={'hour': 0})

However, the block_staircase function makes use of the history function to reconstruct staircase series. For this reason, the response time may be a bit long depending on the time span of the staircase series and the number of insertions.

To optimize the total execution time, we could retrieve the history of "my_forecast" series, store it in memory and reconstruct the staircase series from it. This can be done using the historycache object from the tshsitory.tsio module, as follows

 >>> from tshistory.tsio import historycache
 >>>
 >>> hist = tsa.history('my_forecast'
                        from_value_date=pd.Timestamp('2020-01-01', tz="utc"),
                        to_value_date=pd.Timestamp('2022-01-01', tz="utc"))
 >>> hcache = historycache('my_forecast_cache', hist=hist, tzaware=True)
 >>>
 >>> ts_1da = hcache.block_staircase(from_value_date=pd.Timestamp('2020-01-01', tz="utc"),
                                     to_value_date=pd.Timestamp('2022-01-01', tz="utc"),
                                     revision_freq={'days': 1},
                                     revision_time={'hour': 9},
                                     revision_tz='utc',
                                     maturity_offset={'days': 1},
                                     maturity_time={'hour': 0})
 >>> ts_2da = hcache.block_staircase(from_value_date=pd.Timestamp('2020-01-01', tz="utc"),
                                     to_value_date=pd.Timestamp('2022-01-01', tz="utc"),
                                     revision_freq={'days': 1},
                                     revision_time={'hour': 9},
                                     revision_tz='utc',
                                     maturity_offset={'days': 2},
                                     maturity_time={'hour': 0})

In the example above, the history function of tshistory API is called once. Then the staircase series are computed using the history data stored in memory using the historycache object, which avoids one API call and reduce execution time.

This example assumes the series "my_forecast" is timezone-aware. In the case of timezone-naive series, the tzaware parameter of historycache should be adapted accordingly.

The historycache class also provides a staircase method, so this technique can also be used for basic staircase computation.

The API object

In the few examples above we manipulate the time series through an object that talks directly to the postgresql back end.

It is possible to also talk to a rest api using the same api, like shown below and proceed exactly like in the above code examples:

 >>> from tshistory.api import timeseries
 >>>
 >>> tsa = timeseries('http://my.timeseries.info/api')

Using an HTTP/REST end point

For the rest api, you need to build a small flask app like this (in an app.py module):

from flask import Flask

from tshistory.api import timeseries
from tshistory.http.server import blueprint as blueprint


def make_app(dburi):
    app = Flask('my-timeseries-app')
    app.register_blueprint(
        blueprint(timeseries(dburi)),
        url_prefix='/api'
    )
    return app

Then, you can start it in development mode like this:

app = make_app('postgresql://me:password@localhost/mydb')
app.run('192.168.1.1', 8080)

or just leave it to a wsgi container in e.g. a wsgi.py module:

from my_series_app.app import make_app

app = make_app('postgresql://me:password@localhost/mydb')

API surface

For now we only provide a list of supported methods.

Information access (read methods)

• catalog

• exists

• get

• history

• interval

• metadata

• staircase

• block_staircase

• type

Information update (write methods)

• update

• update_metadata

• replace

• rename

• delete

Command line

Basic operations

A command line tool is provided, called tsh. It provides its usage guidelines:

 $ tsh
 Usage: tsh [OPTIONS] COMMAND [ARGS]...

 Options:
   --help  Show this message and exit.

Commands:
  check    coherence checks of the db
  get      show a serie in its current state
  history  show a serie full history
  info     show global statistics of the repository
  init-db  initialize an new db.
  log      show revision history of entire repository or...
  view     visualize time series through the web

Info provides an overview of the time series repository (number of committed changes, number and series and their names).

 $ tsh info postgres://babar:babarpassword@dataserver:5432/banana_studies
 changeset count: 209
 series count:    144
 series names:    banana_spot_price, banana_trades, banana_turnover

Log provides the full history of editions to time series in the repository.

 $ tsh log postgres://babar:babar@dataserver:5432/banana_studies --limit 3
 revision: 206
 author:   BABAR
 date:     2017-06-06 15:32:51.502507
 series:   banana_spot_price

 revision: 207
 author:   BABAR
 date:     2017-06-06 15:32:51.676507
 series:   banana_trades

 revision: 209
 author:   CELESTE
 date:     2017-06-06 15:32:51.977507
 series:   banana_turnover

All options of all commands can be obtained by using the --help switch:

 $ tsh log --help
 Usage: tsh log [OPTIONS] DB_URI

 Options:
   -l, --limit TEXT
   --show-diff
   -s, --serie TEXT
   --from-rev TEXT
   --to-rev TEXT
   --help            Show this message and exit.

Extensions

It is possible to augment the tsh command with new subcommands (or augment, modify existing commands).

Any program doing so must define a new command and declare a setup tools entry point named tshistory:subcommand as in e.g.:

    entry_points={'tshistory.subcommands': [
        'view=tsview.command:view'
    ]}

For instance, the tsview python package provides such a view subcommand for generic time series visualisation.

Status

It is currently considered beta software even though it has been in production for two years. It is still evolving. Schema/Database changes come with migration procedure using the tsh utility.

When it is good: if you do mostly appends (and occasional edits in the past) it will store data in a very compact way.

Reading any version of the series will always be the fastest (io-bound) operation.

Alternative backend storage and storage strategies will be considered in the future.