An intuitive library tracking dates and timeseries in common using NumPy arrays.

## Project description

# Thymus-timeseries

An intuitive library tracking dates and timeseries in common using numpy

arrays.

When working with arrays of timeseries, the manipulation process can easily

cause mismatching sets of arrays in time, arrays in the wrong order, slow down

the analysis, and lead to generally spending more time to ensure consistency.

This library attempts to address the problem in a way that enables ready access

to the current date range, but stays out of your way most of the time.

Essentially, this library is a wrapper around numpy arrays.

This library grew out of the use of market and trading data. The

timeseries is typically composed of regular intervals but with gaps

such as weekends and holidays. In the case of intra-day data, there are

interuptions due to periods when the market is closed or gaps in trading.

While the library grew from addressing issues associated with market

data, the implementation does not preclude use in other venues. Direct

access to the numpy arrays is expected and the point of being able to use the

library.

## Dependencies

Other than NumPy being installed, there are no other requirements.

## Installation

pip install thymus-timeseries

## A Brief Look at Capabilities.

### Creating a Small Sample Timeseries Object

As a first look, we will create a small timeseries object and show a few ways

that it can used. For this example, we will use daily data.

```

from datetime import datetime

import numpy as np

from thymus.timeseries import Timeseries

ts = Timeseries()

# elements of Timeseries()

key: (an optional identifier for the timeseries)

columns: [] (an optional list of column names for the data)

frequency: d (the d in this case refers to the default daily data.

current frequencies supported are sec, min, h, d, w,

m, q, y)

dseries: (this is a numpy array of dates in numeric format)

tseries: (this is a numpy array of data. most of the work takes

place here.)

end-of-period: True (this is a default indicating that the data is as of

the end of the data. This only comes into play when

converting from one frequency to another and will

be ignored for the moment.)

```

While normal usage of the timeseries object would involve pulling data from a

database and inserting data into the timeseries object, we will use a

quick-and-dirty method of inputting some data. Dates are stored as either

ordinals or timestamps, avoiding clogging up memory with large sets of datetime

objects. Because it is daily data, ordinals will be used for this example.

```

ts = Timeseries()

start_date = datetime(2015, 12, 31).toordinal()

ts.dseries = start_date + np.arange(10)

ts.tseries = np.arange(10)

ts.make_arrays()

```

We created an initial timeseries object. It starts at the end of

2015 and continues for 10 days. Setting the values in **dseries** and

**tseries**

can be somewhat sloppy. For example, a list could be assigned initially to

either **dseries** (the dates) and a numpy array to **tseries** (the values).

The use of the **make_arrays()** function converts the date series to an int32

array (because they are ordinal values) and **tseries** to a float64 array. The

idea is that the data might often enter the timeseries object as lists, but

then be converted to arrays of appropriate format for use.

The completed timeseries object is:

```

print(ts)

<Timeseries>

key:

columns: []

frequency: d

daterange: ('2015-12-31', '2016-01-09')

end-of-period: True

shape: (10,)

```

You can see the date range contained in the date series. The shape refers

to the shape of the **tseries** array. **key** and **columns** are free-form,

available to update as appropriate to identify the timeseries and content of

the columns. Again, the **end-of-period** flag can be ignored right now.

## Selection

Selection of elements is the same as numpy arrays. Currently, our sample has

10 elements.

```

print(ts[:5])

<Timeseries>

key:

columns: []

frequency: d

daterange: ('2015-12-31', '2016-01-04')

end-of-period: True

shape: (5,)

```

Note how the date range above reflects the selected elements.

```

ts1 = ts % 2 == 0

ts1.tseries

[ True False True False True False True False True False]

```

We can isolate the dates of even numbers:

```

# note that tseries, not the timeseries obj, is explicitly used with

# np.argwhere. More on when to operate directly on tseries later.

evens = np.argwhere((ts % 2 == 0).tseries)

ts_even = ts[evens]

# this just prints a list of date and value pairs only useful with

# very small sets (or examples like this)

print(ts_even.items('str'))

('2015-12-31', '[0.0]')

('2016-01-02', '[2.0]')

('2016-01-04', '[4.0]')

('2016-01-06', '[6.0]')

('2016-01-08', '[8.0]')

```

## Date-based Selection

So let us use a slightly larger timeseries. 1000 rows 2 columns of data. And,

use random values to ensure uselessness.

```

ts = Timeseries()

start_date = datetime(2015, 12, 31).toordinal()

ts.dseries = start_date + np.arange(1000)

ts.tseries = np.random.random((1000, 2))

ts.make_arrays()

print(ts)

<Timeseries>

key:

columns: []

frequency: d

daterange: ('2015-12-31', '2018-09-25')

end-of-period: True

shape: (1000, 2)

```

You can select on the basis of date ranges, but first we will use a row number

technique that is based on slicing. This function is called **trunc()** for

truncation.

```

# normal truncation -- you will end up with a timeseries with row 100

# through 499. This provides in-place execution.

ts.trunc(start=100, finish=500)

# this version returns a new timeseries, effective for chaining.

ts1 = ts.trunc(start=100, finish=500, new=True)

```

But suppose you want to select a specific date range? This leads to the next

function, **truncdate()**.

```

# select using datetime objects

ts1 = ts.truncdate(

start=datetime(2017, 1, 1),

finish=datetime(2017, 12, 31),

new=True)

print(ts1)

<Timeseries>

key:

columns: []

frequency: d

daterange: ('2017-01-01', '2017-12-31')

end-of-period: True

shape: (365, 2)

```

As you might expect, the timeseries object has a date range of all the days

during 2017. But see how this is slightly different than slicing. When you use

**truncdate()** it selects everything within the date range inclusive of the

ending date as well. The idea is to avoid having to always find one day after

the date range that you want to select to accommodate slicing behavior. This

way is more convenient.

You can also convert data from a higer frequency to a lower frequency. Suppose

we needed monthly data for 2017 from our timeseries.

```

start = datetime(2017, 1, 1)

finish = datetime(2017, 12, 31)

ts1 = ts.truncdate(start=start, finish=finish, new=True).convert('m')

print(ts1.items('str'))

('2017-01-31', '[0.1724835781570483, 0.9856812220255055]')

('2017-02-28', '[0.3855043513164875, 0.30697511661843124]')

('2017-03-31', '[0.7067982987769881, 0.7680886691626396]')

('2017-04-30', '[0.07770763295126926, 0.04697651222041588]')

('2017-05-31', '[0.4473657194650975, 0.49443624153533783]')

('2017-06-30', '[0.3793816656495891, 0.03646544387811124]')

('2017-07-31', '[0.2783335012003322, 0.5144979569785825]')

('2017-08-31', '[0.9261879195281345, 0.6980224313957553]')

('2017-09-30', '[0.09531834159018227, 0.5435208082899813]')

('2017-10-31', '[0.6865842769906441, 0.7951735180348887]')

('2017-11-30', '[0.34901775001111657, 0.7014208950555662]')

('2017-12-31', '[0.4731393617405252, 0.630488855197775]')

```

Or yearly. In this case, we use a flag that governs whether to include the partial period

leading up to the last year. The default includes it. However, when unwanted the flag,

**include_partial** can be set to False.

```

ts1 = ts.convert('y', include_partial=True)

print(ts1.items('str'))

('2015-12-31', '[0.2288539210230056, 0.288320541664724]')

('2016-12-31', '[0.5116274142615629, 0.21680312154651182]')

('2017-12-31', '[0.4731393617405252, 0.630488855197775]')

('2018-09-25', '[0.7634145837512148, 0.32026411425902257]')

ts2 = ts.convert('y', include_partial=False)

print(ts2.items('str'))

('2015-12-31', '[[0.2288539210230056, 0.288320541664724]]')

('2016-12-31', '[[0.5116274142615629, 0.21680312154651182]]')

('2017-12-31', '[[0.4731393617405252, 0.630488855197775]]')

```

## Combining Timeseries

Suppose you want to combine multiple timeseries together that are of different

lengths? In this case we assume that the two timeseries end on the same date,

but one has a longer tail than the other. However, the operation that you need

requires common dates.

By **combine** we mean instead of two timeseries make one timeseries that has

the columns of both.

```

ts_short = Timeseries()

ts_long = Timeseries()

end_date = datetime(2016, 12, 31)

ts_short.dseries = [

(end_date + timedelta(days=-i)).toordinal()

for i in range(5)]

ts_long.dseries = [

(end_date + timedelta(days=-i)).toordinal()

for i in range(10)]

ts_short.tseries = np.zeros((5))

ts_long.tseries = np.ones((10))

ts_short.make_arrays()

ts_long.make_arrays()

ts_combine = ts_short.combine(ts_long)

print(ts.items('str'))

('2016-12-31', '[0.0, 1.0]')

('2016-12-30', '[0.0, 1.0]')

('2016-12-29', '[0.0, 1.0]')

('2016-12-28', '[0.0, 1.0]')

('2016-12-27', '[0.0, 1.0]')

```

The combine function has a couple variations. While it can be helpful to automatically discard the

unwanted rows, you can also enforce that combining does not take place if the number of rows do not

match. Also, you can build out the missing information with padding to create a timeseries that has

the length of the longest timeseries.

```

# this would raise an error -- the two are different lengths

ts_combine = ts_short.combine(ts_long discard=False)

# this combines, and fills 99 as a missing value

ts_combine = ts_short.combine(ts_long discard=False, pad=99)

print(ts_combine.items('str'))

('2016-12-31', '[0.0, 1.0]')

('2016-12-30', '[0.0, 1.0]')

('2016-12-29', '[0.0, 1.0]')

('2016-12-28', '[0.0, 1.0]')

('2016-12-27', '[0.0, 1.0]')

('2016-12-26', '[99.0, 1.0]')

('2016-12-25', '[99.0, 1.0]')

('2016-12-24', '[99.0, 1.0]')

('2016-12-23', '[99.0, 1.0]')

('2016-12-22', '[99.0, 1.0]')

```

The combining can also receive multiple timeseries.

```

ts_combine = ts_short.combine([ts_long, ts_long, ts_long])

print(ts_combine.items('str'))

('2016-12-31', '[0.0, 1.0, 1.0, 1.0]')

('2016-12-30', '[0.0, 1.0, 1.0, 1.0]')

('2016-12-29', '[0.0, 1.0, 1.0, 1.0]')

('2016-12-28', '[0.0, 1.0, 1.0, 1.0]')

('2016-12-27', '[0.0, 1.0, 1.0, 1.0]')

```

## Splitting Timeseries

In some ways it would make sense to mirror the **combine()** function

with a **split()** from an aesthetic standpoint. However, splitting is very

straight-forward without such a function. For example, suppose you want a

timeseries that only has the the first two columns from our previous example.

As you can see in the ts_split tseries, the first two columns were taken.

```

ts_split = ts_combine[:, :2]

print(ts_split.items('str'))

('2016-12-31', '[0.0, 1.0]')

('2016-12-30', '[0.0, 1.0]')

('2016-12-29', '[0.0, 1.0]')

('2016-12-28', '[0.0, 1.0]')

('2016-12-27', '[0.0, 1.0]')

```

## Arithmetic Operations

We have combined timeseries together to stack up rows in common. In

addition, we looked at the issue of mismatched lengths. Now we will look at

arithmetic approaches and some of the design decisions and tradeoffs associated

with mathematical operations.

We will start with the **add()** function. First, if we assume that all we are

adding together are arrays that have exactly the same dateseries, and

therefore the same length, and we assume they have exactly the same number of

columns, then the whole question becomes trivial. If we relax those

constraints, then some choices need to be made.

We will use the long and short timeseries from the previous example.

```

# this will fail due to dissimilar lengths

ts_added = ts_short.add(ts_long, match=True)

# this will work

ts_added = ts_short.add(ts_long, match=False)

[ 1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]

```

The **add()** function checks to see if the number of columns match. If they do

not an error is raised. If the **match** flag is True, then it also checks

that all the dates in both timeseries match prior to the operation.

If **match** is False, then as long as the columns are compatible, the

operation can take place. It also supports the concept of sparse arrays as

well. For example, suppose you have a timeseries that is primary, but you would

like to add in a timeseries values from only a few dates within the range. This

function will find the appropriate dates adding in the values at just those

rows.

To summarize, all dates in common to both timeseries will be included in the

new timeseries if **match** is False.

Because the previous function is somewhat specialized, you can assume that the

checking of common dates and creating the new timeseries can be somewhat slower

than other approaches.

If we assume some commonalities about our timeseries, then we can do our work

in a more intuitive fashion.

### Assumptions of Commonality

Let us assume that our timeseries might be varying in length, but we absolutely

know what either our starting date or ending date is. And, let us assume that

all the dates for the periods in common to the timeseries match.

If we accept those assumptions, then a number of operations become quite easy.

The timeseries object can accept simple arithmetic as if it is an array. It

automatically passes the values on to the **tseries** array. If the two arrays

are not the same length the longer array is truncated to the shorter length. So

if you were add two arrays together that end at the same date, you would want

to sort them latest date to earliest date using the function

**sort_by_date()**.

### Examples

```

# starting tseries

ts.tseries

[ 0. 1. 2. 3. 4. 5. 6. 7. 8. 9.]

(ts + 3).tseries

[ 3. 4. 5. 6. 7. 8. 9. 10. 11. 12.]

# Also, reverse (__radd__)

(3 + ts).tseries

[ 3. 4. 5. 6. 7. 8. 9. 10. 11. 12.]

# of course not just addition

5 * ts.tseries

[ 0. 5. 10. 15. 20. 25. 30. 35. 40. 45.]

```

Also, in-place operations. But first, we will make a copy.

```

ts1 = ts.clone()

ts1.tseries /= 3

print(ts1.tseries)

[0.0

0.3333333333333333

0.6666666666666666

1.0

1.3333333333333333

1.6666666666666667

2.0

2.3333333333333335

2.6666666666666665

3.0]

ts1 = ts ** 3

ts1.tseries

0.0

1.0

8.0

27.0

64.0

125.0

216.0

343.0

512.0

729.0

ts1 = 10 ** ts

ts1.tseries

[1.0

10.0

100.0

1000.0

10000.0

100000.0

1000000.0

10000000.0

100000000.0

1000000000.0]

```

In other words, the normal container functions you can use with numpy arrays

are available to the timeseries objects. The following container functions for

arrays are supported.

```

__pow__ __add__ __rsub__ __sub__ __eq__ __ge__ __gt__ __le__

__lt__ __mod__ __mul__ __ne__ __radd__ __rmod__ __rmul__ __rpow__

__abs__ __pos__ __neg__ __invert__ __rdivmod__ __rfloordiv__

__floordiv__ __truediv__

__rtruediv__ __divmod__

__and__ __or__ __ror__ __rand__ __rxor__ __xor__ __rshift__

__rlshift__ __lshift__ __rrshift__

__iadd__ __ifloordiv__ __imod__ __imul__ __ipow__ __isub__

__itruediv__]

__iand__ __ilshift__ __ior__ __irshift__ __ixor__

```

### Functions of Arrays Not Supported

The purpose the timeseries objects is to implement an intuitive usage of

timeseries objects in a fashion that is consistent with NumPy. However, it is

not intended to replace functions that are better handled explicitly with

the **dseries** and **tseries** arrays directly. The difference will be clear

by

comparing the list of functions for the timeseries object versus a numpy array. Most of the

functions of the timeseries object is related to handling the commonality of date series with

time series. You can see that the bulk of the thymus functions relate to maintenance of the

coordination betwee the date series and timeseries. The meat of the functions still lie with the

numpy arrays.

```

# timeseries members and functions:

ts.add ts.daterange ts.get_pcdiffs ts.series_direction

ts.as_dict ts.datetime_series ts.header ts.set_ones

ts.as_json ts.dseries ts.if_dseries_match ts.set_zeros

ts.as_list ts.end_date ts.if_tseries_match ts.shape

ts.clone ts.end_of_period ts.items ts.sort_by_date

ts.closest_date ts.extend ts.key ts.start_date

ts.columns ts.fmt_date ts.lengths ts.trunc

ts.combine ts.frequency ts.make_arrays ts.truncdate

ts.common_length ts.get_date_series_type ts.months ts.tseries

ts.convert ts.get_datetime ts.replace ts.years

ts.date_native ts.get_diffs ts.reverse

ts.date_string_series ts.get_duped_dates ts.row_no

# numpy functions in the arrays

ts.tseries.T ts.tseries.cumsum ts.tseries.min ts.tseries.shape

ts.tseries.all ts.tseries.data ts.tseries.nbytes ts.tseries.size

ts.tseries.any ts.tseries.diagonal ts.tseries.ndim ts.tseries.sort

ts.tseries.argmax ts.tseries.dot ts.tseries.newbyteorder ts.tseries.squeeze

ts.tseries.argmin ts.tseries.dtype ts.tseries.nonzero ts.tseries.std

ts.tseries.argpartition ts.tseries.dump ts.tseries.partition ts.tseries.strides

ts.tseries.argsort ts.tseries.dumps ts.tseries.prod ts.tseries.sum

ts.tseries.astype ts.tseries.fill ts.tseries.ptp ts.tseries.swapaxes

ts.tseries.base ts.tseries.flags ts.tseries.put ts.tseries.take

ts.tseries.byteswap ts.tseries.flat ts.tseries.ravel ts.tseries.tobytes

ts.tseries.choose ts.tseries.flatten ts.tseries.real ts.tseries.tofile

ts.tseries.clip ts.tseries.getfield ts.tseries.repeat ts.tseries.tolist

ts.tseries.compress ts.tseries.imag ts.tseries.reshape ts.tseries.tostring

ts.tseries.conj ts.tseries.item ts.tseries.resize ts.tseries.trace

ts.tseries.conjugate ts.tseries.itemset ts.tseries.round ts.tseries.transpose

ts.tseries.copy ts.tseries.itemsize ts.tseries.searchsorted ts.tseries.var

ts.tseries.ctypes ts.tseries.max ts.tseries.setfield ts.tseries.view

ts.tseries.cumprod ts.tseries.mean ts.tseries.setflags

```

### Other Date Functions

Variations on a theme:

```

# truncation

ts.truncdate(

start=datetime(2017, 1, 1),

finish=datetime(2017, 12, 31))

# just start date etc.

ts.truncdate(

start=datetime(2017, 1, 1))

# this was in date order but suppose it was in reverse order?

# this result will give the same answer

ts1 = ts.truncdate(

start=datetime(2017, 1, 1),

new=True)

ts.reverse()

ts1 = ts.truncdate(

start=datetime(2017, 1, 1),

new=True)

# use the date format native to the dateseries (ordinal / timestamp)

ts1 = ts.truncdate(

start=datetime(2017, 1, 1).toordinal(),

new=True)

# suppose you start with a variable that represents a date range

# date range can be either a list or tuple

ts.truncdate(

[datetime(2017, 1, 1), datetime(2017, 12, 31)])

```

## Assorted Date Functions

```

# native format

ts.daterange()

(735963, 735972)

# str format

ts.daterange('str')

('2015-12-31', '2016-01-09')

# datetime format

ts.daterange('datetime')

(datetime.datetime(2015, 12, 31, 0, 0), datetime.datetime(2016, 1, 9, 0, 0))

# native format

ts.start_date(); ts.end_date()

735963 735972

# str format

ts.start_date('str'); ts.end_date('str')

2015-12-31 2016-01-09

# datetime format

ts.start_date('datetime'); ts.end_date('datetime')

2015-12-31 00:00:00 2016-01-09 00:00:00

```

Sometimes it is helpful to find a particular row based on the date. Also, that date might not be in

the dateseries, and so, the closest date will suffice.

We will create a sample timeseries to illustrate.

```

ts = Timeseries()

ts.dseries = []

ts.tseries = []

start_date = datetime(2015, 12, 31)

for i in range(40):

date = start_date + timedelta(days=i)

if date.weekday() not in [5, 6]: # skipping weekends

ts.dseries.append(date.toordinal())

ts.tseries.append(i)

ts.make_arrays()

# row_no, date

(0, '2015-12-31')

(1, '2016-01-01')

(2, '2016-01-04')

(3, '2016-01-05')

(4, '2016-01-06')

(5, '2016-01-07')

(6, '2016-01-08')

(7, '2016-01-11')

(8, '2016-01-12')

(9, '2016-01-13')

(10, '2016-01-14')

(11, '2016-01-15')

(12, '2016-01-18')

(13, '2016-01-19')

(14, '2016-01-20')

(15, '2016-01-21')

(16, '2016-01-22')

(17, '2016-01-25')

(18, '2016-01-26')

(19, '2016-01-27')

(20, '2016-01-28')

(21, '2016-01-29')

(22, '2016-02-01')

(23, '2016-02-02')

(24, '2016-02-03')

(25, '2016-02-04')

(26, '2016-02-05')

(27, '2016-02-08')

date1 = datetime(2016, 1, 7) # existing date within date series

date2 = datetime(2016, 1, 16) # date falling on a weekend

date3 = datetime(2015, 6, 16) # date prior to start of date series

date4 = datetime(2016, 3, 8) # date after to end of date series

# as datetime and in the series

existing_row = ts.row_no(rowdate=date1, closest=1)

5

existing_date = ts.closest_date(rowdate=date1, closest=1)

print(datetime.fromordinal(existing_date))

2016-01-07 00:00:00

# as datetime but date not in series

next_row = ts.row_no(rowdate=date2, closest=1)

12

next_date = ts.closest_date(rowdate=date2, closest=1)

print(datetime.fromordinal(next_date))

2016-01-18 00:00:00

prev_row = ts.row_no(rowdate=date2, closest=-1)

11

prev_date = ts.closest_date(rowdate=date2, closest=-1)

print(datetime.fromordinal(prev_date))

2016-01-15 00:00:00

# this will fail -- date is outside the date series

# as datetime but date not in series, look for earlier date

ts.closest_date(rowdate=date3, closest=-1)

# this will fail -- date is outside the date series

ts.closest_date(rowdate=date4, closest=1)

```

## Functions by Category

### Output

#### ts.as_dict()

Returns the time series as a dict with the date as the key and without

the header information.

#### ts.as_json(indent=2)

This function returns the timeseries in JSON format and includes the

header information.

#### ts.as_list()

Returns the timeseries as a list.

#### ts.header()

This function returns a dict of the non-timeseries data.

#### ts.items(fmt=None)

This function returns the date series and the time series as if it

is in one list. The term items used to suggest the iteration of dicts

where items are the key, value combination.

if fmt == 'str':

the dates are output as strings

#### ts.months(include_partial=True)

This function provides a quick way to summarize daily (or less)

as monthly data.

It is basically a pass-through to the convert function with more

decoration of the months.

Usage:

months(include_partial=True)

returns a dict with year-month as keys

#### ts.years(include_partial=True)

This function provides a quick way to summarize daily (or less)

as yearly data.

It is basically a pass-through to the convert function with more

decoration of the years.

Usage:

years(include_partial=True)

returns a dict with year as keys

#### ts.datetime_series()

This function returns the dateseries converted to a list of

datetime objects.

#### ts.date_string_series(dt_fmt=None)

This function returns a list of the dates in the timeseries as

strings.

Usage:

self.date_string_series(dt_fmt=None)

dt_fmt is a datetime mask to alter the default formatting.

### Array Manipulation

#### ts.add(ts, match=True)

Adds two timeseries together.

if match is True:

means there should be a one to one corresponding date in each time

series. If not raise error.

else:

means that timeseries with sporadic or missing dates can be added

Note: this does not evaluate whether both timeseries have the same

number of columns. It will fail if they do not.

Returns the timeseries. Not in-place.

#### ts.clone()

This function returns a copy of the timeseries.

#### ts.combine(tss, discard=True, pad=None)

This function combines timeseries into a single array. Combining in

this case means accumulating additional columns of information.

Truncation takes place at the end of rows. So if the timeseries is

sorted from latest dates to earliest dates, the older values would be

removed.

Usage:

self.combine(tss, discard=True, pad=None)

Think of tss as the plural of timeseries.

If discard:

Will truncate all timeseries lengths down to the shortest

timeseries.

if discard is False:

An error will be raised if the all the lengths do not match

unless:

if pad is not None:

the shorter timeseries will be padded with the value pad.

Returns the new ts.

#### ts.common_length(ts1, ts2)

This static method trims the lengths of two timeseries and returns two

timeseries with the same length.

The idea is that in order to do array operations there must be a

common length for each timeseries.

Reflecting the bias for using timeseries sorted from latest info to

earlier info, truncation takes place at the end of the array. That

way older less important values are removed if necessary.

Usage:

ts1_new, ts2_new = self.common_length(ts1, ts2)

#### ts.convert(new_freq, include_partial=True, **kwargs)

This function returns the timeseries converted to another frequency,

such as daily to monthly.

Usage:

convert(new_freq, include_partial=True, **kwargs)

The only kwarg is

weekday=<some value>

This is used when converting to weekly data. The weekday number

corresponds to the the datetime.weekday() function.

#### ts.extend(ts, overlay=True)

This function combines a timeseries to another, taking into account the

possibility of overlap.

This assumes that the frequency is the same.

This function is chiefly envisioned to extend a timeseries with

additional dates.

Usage:

self.extend(ts, overlay=True)

If overlay is True then the incoming timeseries will overlay

any values that are duplicated.

#### ts.trunc(start=None, finish=None, new=False)

This function truncates in place, typically.

truncate from (start:finish)

remember start is lowest number, latest date

This truncation works on the basis of slicing, so

finish is not inclusive.

Usage:

self.trunc(start=None, finish=None, new=False)

#### ts.truncdate(start=None, finish=None, new=False)

This function truncates in place on the basis of dates.

Usage:

self.truncdate(start=None, finish=None, new=False)

start and finish are dates, input as either datetime or the actual

internal format of the **dseries** (ordinals or timestamps).

If the dates are not actually in the list, the starting date will

be the next viable date after the start date requested. If the finish

date is not available, the previous date from the finish date will be

the last.

If new is True, the timeseries will not be modified in place. Rather

a new timeseries will be returned instead.

#### ts.replace(ts, match=True)

This function replaces values where the dates match an incoming

timeseries. So if the incoming date on the timeseries matches, the

value in the current timeseries will be replaced by the incoming

timeseries.

Usage:

self.replace(ts, match=True)

If match is False, the incoming timseries may have dates not found in

the self timeseries.

Returns the modified timeseries. Not in place.

#### ts.reverse()

This function does in-place reversal of the timeseries and dateseries.

#### ts.get_diffs()

This function gets the differences between values from date to date in

the timeseries.

#### ts.get_pcdiffs()

This function gets the percent differences between values in the

timeseries.

No provision for dividing by zero here.

#### ts.set_ones(fmt=None, new=False)

This function converts an existing timeseries to ones using the same

shape as the existing timeseries.

It is used as a convenience to create an empty timeseries with a

specified date range.

if fmt use as shape

usage:

set_ones(self, fmt=None, new=False)

#### ts.set_zeros(fmt=None, new=False)

This function converts an existing timeseries to zeros using the same

shape as the existing timeseries.

It is used as a convenience to create an empty timeseries with a

specified date range.

if fmt use as shape

usage:

set_zeros(self, fmt=None, new=False)

#### ts.sort_by_date(reverse=False, force=False)

This function converts a timeseries to either date order or reverse

date order.

Usage:

sort_by_date(self, reverse=False, force=False)

If reverse is True, then order will be newest to oldest.

If force is False, the assumption is made that comparing the first

and last date will determine the current order of the timeseries. That

would mean that unnecessary sorting can be avoided. Also, if the order

needs to be reversed, the sort is changed via the less expensive

reverse function.

If dates and values are in no particular order, with force=True, the

actual sort takes place.

This function changes the data in-place.

### Evaluation

#### ts.daterange(fmt=None)

This function returns the starting and ending dates of the timeseries.

Usage:

self.daterange()

(735963, 735972)

self.daterange('str')

('2015-12-31', '2016-01-09')

self.daterange('datetime')

(datetime(2015, 12, 31, 0, 0),

datetime.datetime(2016, 1, 9, 0, 0))

#### ts.start_date(fmt=None)

This function returns the starting date of the timeseries in its

native value, timestamp or ordinal.

If fmt is 'str' returns in string format

If fmt is 'datetime' returns in string format

#### ts.end_date(fmt=None)

This funtcion returns the ending date of the timeseries in its native

value, timestamp or ordinal.

If fmt is 'str' returns in string format

If fmt is 'datetime' returns in string format

#### ts.get_duped_dates()

This function pulls dates that are duplicated. This is to be used to

locate timeseries that are faulty.

Usage:

get_duped_dates()

returns [[odate1, count], [odate2, count]]

#### ts.series_direction()

if a lower row is a lower date, then 1 for ascending

if a lower row is a higher date then -1 for descending

#### ts.get_date_series_type()

This function returns the date series type associated with the

timeseries. The choices are TS_ORDINAL or TS_TIMESTAMP.

#### ts.if_dseries_match(ts)

This function returns True if the date series are the same.

#### ts.if_tseries_match(ts)

This function returns True if the time series are the same.

### Utilities

#### ts.date_native(date)

This awkwardly named function returns a date in the native format of

of the timeseries, namely ordinal or timestamp.

#### ts.row_no(rowdate, closest=0, no_error=False)

Shows the row in the timeseries

Usage:

ts.row(rowdate=<datetime>)

ts.row(rowdate=<date as either ordinal or timestamp>)

Returns an error if the date is not found in the index

if closest is invoked:

closest = 1

find the closest date after the rowdate

closest = -1

find the closest date before the rowdate

If no_error

returns -1 instead of raising an error if the date was

outside of the timeseries.

#### ts.get_datetime(date)

This function returns a date as a datetime object.

This takes into account the type of date stored in **dseries**.

Usage:

self.get_datetime(date)

#### ts.lengths()

This function returns the lengths of both the date series and time

series. Both numbers are included in case a mismatch has occurred.

#### ts.shape()

This function return the shape of the timeseries. This is a shortcut

to putting in ts.tseries.shape.

#### ts.fmt_date(numericdate, dt_type, dt_fmt=None)

This static method accepts a date and converts it to

the format used in the timeseries.

#### ts.make_arrays()

Convert the date and time series lists (if so) to numpy arrays

#### ts.get_fromDB(**kwargs)

This is just a stub to suggest a viable name for getting data from a

database.

#### ts.save_toDB(**kwargs):

This is just a stub to suggest a viable name for saving data to a

database.

An intuitive library tracking dates and timeseries in common using numpy

arrays.

When working with arrays of timeseries, the manipulation process can easily

cause mismatching sets of arrays in time, arrays in the wrong order, slow down

the analysis, and lead to generally spending more time to ensure consistency.

This library attempts to address the problem in a way that enables ready access

to the current date range, but stays out of your way most of the time.

Essentially, this library is a wrapper around numpy arrays.

This library grew out of the use of market and trading data. The

timeseries is typically composed of regular intervals but with gaps

such as weekends and holidays. In the case of intra-day data, there are

interuptions due to periods when the market is closed or gaps in trading.

While the library grew from addressing issues associated with market

data, the implementation does not preclude use in other venues. Direct

access to the numpy arrays is expected and the point of being able to use the

library.

## Dependencies

Other than NumPy being installed, there are no other requirements.

## Installation

pip install thymus-timeseries

## A Brief Look at Capabilities.

### Creating a Small Sample Timeseries Object

As a first look, we will create a small timeseries object and show a few ways

that it can used. For this example, we will use daily data.

```

from datetime import datetime

import numpy as np

from thymus.timeseries import Timeseries

ts = Timeseries()

# elements of Timeseries()

key: (an optional identifier for the timeseries)

columns: [] (an optional list of column names for the data)

frequency: d (the d in this case refers to the default daily data.

current frequencies supported are sec, min, h, d, w,

m, q, y)

dseries: (this is a numpy array of dates in numeric format)

tseries: (this is a numpy array of data. most of the work takes

place here.)

end-of-period: True (this is a default indicating that the data is as of

the end of the data. This only comes into play when

converting from one frequency to another and will

be ignored for the moment.)

```

While normal usage of the timeseries object would involve pulling data from a

database and inserting data into the timeseries object, we will use a

quick-and-dirty method of inputting some data. Dates are stored as either

ordinals or timestamps, avoiding clogging up memory with large sets of datetime

objects. Because it is daily data, ordinals will be used for this example.

```

ts = Timeseries()

start_date = datetime(2015, 12, 31).toordinal()

ts.dseries = start_date + np.arange(10)

ts.tseries = np.arange(10)

ts.make_arrays()

```

We created an initial timeseries object. It starts at the end of

2015 and continues for 10 days. Setting the values in **dseries** and

**tseries**

can be somewhat sloppy. For example, a list could be assigned initially to

either **dseries** (the dates) and a numpy array to **tseries** (the values).

The use of the **make_arrays()** function converts the date series to an int32

array (because they are ordinal values) and **tseries** to a float64 array. The

idea is that the data might often enter the timeseries object as lists, but

then be converted to arrays of appropriate format for use.

The completed timeseries object is:

```

print(ts)

<Timeseries>

key:

columns: []

frequency: d

daterange: ('2015-12-31', '2016-01-09')

end-of-period: True

shape: (10,)

```

You can see the date range contained in the date series. The shape refers

to the shape of the **tseries** array. **key** and **columns** are free-form,

available to update as appropriate to identify the timeseries and content of

the columns. Again, the **end-of-period** flag can be ignored right now.

## Selection

Selection of elements is the same as numpy arrays. Currently, our sample has

10 elements.

```

print(ts[:5])

<Timeseries>

key:

columns: []

frequency: d

daterange: ('2015-12-31', '2016-01-04')

end-of-period: True

shape: (5,)

```

Note how the date range above reflects the selected elements.

```

ts1 = ts % 2 == 0

ts1.tseries

[ True False True False True False True False True False]

```

We can isolate the dates of even numbers:

```

# note that tseries, not the timeseries obj, is explicitly used with

# np.argwhere. More on when to operate directly on tseries later.

evens = np.argwhere((ts % 2 == 0).tseries)

ts_even = ts[evens]

# this just prints a list of date and value pairs only useful with

# very small sets (or examples like this)

print(ts_even.items('str'))

('2015-12-31', '[0.0]')

('2016-01-02', '[2.0]')

('2016-01-04', '[4.0]')

('2016-01-06', '[6.0]')

('2016-01-08', '[8.0]')

```

## Date-based Selection

So let us use a slightly larger timeseries. 1000 rows 2 columns of data. And,

use random values to ensure uselessness.

```

ts = Timeseries()

start_date = datetime(2015, 12, 31).toordinal()

ts.dseries = start_date + np.arange(1000)

ts.tseries = np.random.random((1000, 2))

ts.make_arrays()

print(ts)

<Timeseries>

key:

columns: []

frequency: d

daterange: ('2015-12-31', '2018-09-25')

end-of-period: True

shape: (1000, 2)

```

You can select on the basis of date ranges, but first we will use a row number

technique that is based on slicing. This function is called **trunc()** for

truncation.

```

# normal truncation -- you will end up with a timeseries with row 100

# through 499. This provides in-place execution.

ts.trunc(start=100, finish=500)

# this version returns a new timeseries, effective for chaining.

ts1 = ts.trunc(start=100, finish=500, new=True)

```

But suppose you want to select a specific date range? This leads to the next

function, **truncdate()**.

```

# select using datetime objects

ts1 = ts.truncdate(

start=datetime(2017, 1, 1),

finish=datetime(2017, 12, 31),

new=True)

print(ts1)

<Timeseries>

key:

columns: []

frequency: d

daterange: ('2017-01-01', '2017-12-31')

end-of-period: True

shape: (365, 2)

```

As you might expect, the timeseries object has a date range of all the days

during 2017. But see how this is slightly different than slicing. When you use

**truncdate()** it selects everything within the date range inclusive of the

ending date as well. The idea is to avoid having to always find one day after

the date range that you want to select to accommodate slicing behavior. This

way is more convenient.

You can also convert data from a higer frequency to a lower frequency. Suppose

we needed monthly data for 2017 from our timeseries.

```

start = datetime(2017, 1, 1)

finish = datetime(2017, 12, 31)

ts1 = ts.truncdate(start=start, finish=finish, new=True).convert('m')

print(ts1.items('str'))

('2017-01-31', '[0.1724835781570483, 0.9856812220255055]')

('2017-02-28', '[0.3855043513164875, 0.30697511661843124]')

('2017-03-31', '[0.7067982987769881, 0.7680886691626396]')

('2017-04-30', '[0.07770763295126926, 0.04697651222041588]')

('2017-05-31', '[0.4473657194650975, 0.49443624153533783]')

('2017-06-30', '[0.3793816656495891, 0.03646544387811124]')

('2017-07-31', '[0.2783335012003322, 0.5144979569785825]')

('2017-08-31', '[0.9261879195281345, 0.6980224313957553]')

('2017-09-30', '[0.09531834159018227, 0.5435208082899813]')

('2017-10-31', '[0.6865842769906441, 0.7951735180348887]')

('2017-11-30', '[0.34901775001111657, 0.7014208950555662]')

('2017-12-31', '[0.4731393617405252, 0.630488855197775]')

```

Or yearly. In this case, we use a flag that governs whether to include the partial period

leading up to the last year. The default includes it. However, when unwanted the flag,

**include_partial** can be set to False.

```

ts1 = ts.convert('y', include_partial=True)

print(ts1.items('str'))

('2015-12-31', '[0.2288539210230056, 0.288320541664724]')

('2016-12-31', '[0.5116274142615629, 0.21680312154651182]')

('2017-12-31', '[0.4731393617405252, 0.630488855197775]')

('2018-09-25', '[0.7634145837512148, 0.32026411425902257]')

ts2 = ts.convert('y', include_partial=False)

print(ts2.items('str'))

('2015-12-31', '[[0.2288539210230056, 0.288320541664724]]')

('2016-12-31', '[[0.5116274142615629, 0.21680312154651182]]')

('2017-12-31', '[[0.4731393617405252, 0.630488855197775]]')

```

## Combining Timeseries

Suppose you want to combine multiple timeseries together that are of different

lengths? In this case we assume that the two timeseries end on the same date,

but one has a longer tail than the other. However, the operation that you need

requires common dates.

By **combine** we mean instead of two timeseries make one timeseries that has

the columns of both.

```

ts_short = Timeseries()

ts_long = Timeseries()

end_date = datetime(2016, 12, 31)

ts_short.dseries = [

(end_date + timedelta(days=-i)).toordinal()

for i in range(5)]

ts_long.dseries = [

(end_date + timedelta(days=-i)).toordinal()

for i in range(10)]

ts_short.tseries = np.zeros((5))

ts_long.tseries = np.ones((10))

ts_short.make_arrays()

ts_long.make_arrays()

ts_combine = ts_short.combine(ts_long)

print(ts.items('str'))

('2016-12-31', '[0.0, 1.0]')

('2016-12-30', '[0.0, 1.0]')

('2016-12-29', '[0.0, 1.0]')

('2016-12-28', '[0.0, 1.0]')

('2016-12-27', '[0.0, 1.0]')

```

The combine function has a couple variations. While it can be helpful to automatically discard the

unwanted rows, you can also enforce that combining does not take place if the number of rows do not

match. Also, you can build out the missing information with padding to create a timeseries that has

the length of the longest timeseries.

```

# this would raise an error -- the two are different lengths

ts_combine = ts_short.combine(ts_long discard=False)

# this combines, and fills 99 as a missing value

ts_combine = ts_short.combine(ts_long discard=False, pad=99)

print(ts_combine.items('str'))

('2016-12-31', '[0.0, 1.0]')

('2016-12-30', '[0.0, 1.0]')

('2016-12-29', '[0.0, 1.0]')

('2016-12-28', '[0.0, 1.0]')

('2016-12-27', '[0.0, 1.0]')

('2016-12-26', '[99.0, 1.0]')

('2016-12-25', '[99.0, 1.0]')

('2016-12-24', '[99.0, 1.0]')

('2016-12-23', '[99.0, 1.0]')

('2016-12-22', '[99.0, 1.0]')

```

The combining can also receive multiple timeseries.

```

ts_combine = ts_short.combine([ts_long, ts_long, ts_long])

print(ts_combine.items('str'))

('2016-12-31', '[0.0, 1.0, 1.0, 1.0]')

('2016-12-30', '[0.0, 1.0, 1.0, 1.0]')

('2016-12-29', '[0.0, 1.0, 1.0, 1.0]')

('2016-12-28', '[0.0, 1.0, 1.0, 1.0]')

('2016-12-27', '[0.0, 1.0, 1.0, 1.0]')

```

## Splitting Timeseries

In some ways it would make sense to mirror the **combine()** function

with a **split()** from an aesthetic standpoint. However, splitting is very

straight-forward without such a function. For example, suppose you want a

timeseries that only has the the first two columns from our previous example.

As you can see in the ts_split tseries, the first two columns were taken.

```

ts_split = ts_combine[:, :2]

print(ts_split.items('str'))

('2016-12-31', '[0.0, 1.0]')

('2016-12-30', '[0.0, 1.0]')

('2016-12-29', '[0.0, 1.0]')

('2016-12-28', '[0.0, 1.0]')

('2016-12-27', '[0.0, 1.0]')

```

## Arithmetic Operations

We have combined timeseries together to stack up rows in common. In

addition, we looked at the issue of mismatched lengths. Now we will look at

arithmetic approaches and some of the design decisions and tradeoffs associated

with mathematical operations.

We will start with the **add()** function. First, if we assume that all we are

adding together are arrays that have exactly the same dateseries, and

therefore the same length, and we assume they have exactly the same number of

columns, then the whole question becomes trivial. If we relax those

constraints, then some choices need to be made.

We will use the long and short timeseries from the previous example.

```

# this will fail due to dissimilar lengths

ts_added = ts_short.add(ts_long, match=True)

# this will work

ts_added = ts_short.add(ts_long, match=False)

[ 1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]

```

The **add()** function checks to see if the number of columns match. If they do

not an error is raised. If the **match** flag is True, then it also checks

that all the dates in both timeseries match prior to the operation.

If **match** is False, then as long as the columns are compatible, the

operation can take place. It also supports the concept of sparse arrays as

well. For example, suppose you have a timeseries that is primary, but you would

like to add in a timeseries values from only a few dates within the range. This

function will find the appropriate dates adding in the values at just those

rows.

To summarize, all dates in common to both timeseries will be included in the

new timeseries if **match** is False.

Because the previous function is somewhat specialized, you can assume that the

checking of common dates and creating the new timeseries can be somewhat slower

than other approaches.

If we assume some commonalities about our timeseries, then we can do our work

in a more intuitive fashion.

### Assumptions of Commonality

Let us assume that our timeseries might be varying in length, but we absolutely

know what either our starting date or ending date is. And, let us assume that

all the dates for the periods in common to the timeseries match.

If we accept those assumptions, then a number of operations become quite easy.

The timeseries object can accept simple arithmetic as if it is an array. It

automatically passes the values on to the **tseries** array. If the two arrays

are not the same length the longer array is truncated to the shorter length. So

if you were add two arrays together that end at the same date, you would want

to sort them latest date to earliest date using the function

**sort_by_date()**.

### Examples

```

# starting tseries

ts.tseries

[ 0. 1. 2. 3. 4. 5. 6. 7. 8. 9.]

(ts + 3).tseries

[ 3. 4. 5. 6. 7. 8. 9. 10. 11. 12.]

# Also, reverse (__radd__)

(3 + ts).tseries

[ 3. 4. 5. 6. 7. 8. 9. 10. 11. 12.]

# of course not just addition

5 * ts.tseries

[ 0. 5. 10. 15. 20. 25. 30. 35. 40. 45.]

```

Also, in-place operations. But first, we will make a copy.

```

ts1 = ts.clone()

ts1.tseries /= 3

print(ts1.tseries)

[0.0

0.3333333333333333

0.6666666666666666

1.0

1.3333333333333333

1.6666666666666667

2.0

2.3333333333333335

2.6666666666666665

3.0]

ts1 = ts ** 3

ts1.tseries

0.0

1.0

8.0

27.0

64.0

125.0

216.0

343.0

512.0

729.0

ts1 = 10 ** ts

ts1.tseries

[1.0

10.0

100.0

1000.0

10000.0

100000.0

1000000.0

10000000.0

100000000.0

1000000000.0]

```

In other words, the normal container functions you can use with numpy arrays

are available to the timeseries objects. The following container functions for

arrays are supported.

```

__pow__ __add__ __rsub__ __sub__ __eq__ __ge__ __gt__ __le__

__lt__ __mod__ __mul__ __ne__ __radd__ __rmod__ __rmul__ __rpow__

__abs__ __pos__ __neg__ __invert__ __rdivmod__ __rfloordiv__

__floordiv__ __truediv__

__rtruediv__ __divmod__

__and__ __or__ __ror__ __rand__ __rxor__ __xor__ __rshift__

__rlshift__ __lshift__ __rrshift__

__iadd__ __ifloordiv__ __imod__ __imul__ __ipow__ __isub__

__itruediv__]

__iand__ __ilshift__ __ior__ __irshift__ __ixor__

```

### Functions of Arrays Not Supported

The purpose the timeseries objects is to implement an intuitive usage of

timeseries objects in a fashion that is consistent with NumPy. However, it is

not intended to replace functions that are better handled explicitly with

the **dseries** and **tseries** arrays directly. The difference will be clear

by

comparing the list of functions for the timeseries object versus a numpy array. Most of the

functions of the timeseries object is related to handling the commonality of date series with

time series. You can see that the bulk of the thymus functions relate to maintenance of the

coordination betwee the date series and timeseries. The meat of the functions still lie with the

numpy arrays.

```

# timeseries members and functions:

ts.add ts.daterange ts.get_pcdiffs ts.series_direction

ts.as_dict ts.datetime_series ts.header ts.set_ones

ts.as_json ts.dseries ts.if_dseries_match ts.set_zeros

ts.as_list ts.end_date ts.if_tseries_match ts.shape

ts.clone ts.end_of_period ts.items ts.sort_by_date

ts.closest_date ts.extend ts.key ts.start_date

ts.columns ts.fmt_date ts.lengths ts.trunc

ts.combine ts.frequency ts.make_arrays ts.truncdate

ts.common_length ts.get_date_series_type ts.months ts.tseries

ts.convert ts.get_datetime ts.replace ts.years

ts.date_native ts.get_diffs ts.reverse

ts.date_string_series ts.get_duped_dates ts.row_no

# numpy functions in the arrays

ts.tseries.T ts.tseries.cumsum ts.tseries.min ts.tseries.shape

ts.tseries.all ts.tseries.data ts.tseries.nbytes ts.tseries.size

ts.tseries.any ts.tseries.diagonal ts.tseries.ndim ts.tseries.sort

ts.tseries.argmax ts.tseries.dot ts.tseries.newbyteorder ts.tseries.squeeze

ts.tseries.argmin ts.tseries.dtype ts.tseries.nonzero ts.tseries.std

ts.tseries.argpartition ts.tseries.dump ts.tseries.partition ts.tseries.strides

ts.tseries.argsort ts.tseries.dumps ts.tseries.prod ts.tseries.sum

ts.tseries.astype ts.tseries.fill ts.tseries.ptp ts.tseries.swapaxes

ts.tseries.base ts.tseries.flags ts.tseries.put ts.tseries.take

ts.tseries.byteswap ts.tseries.flat ts.tseries.ravel ts.tseries.tobytes

ts.tseries.choose ts.tseries.flatten ts.tseries.real ts.tseries.tofile

ts.tseries.clip ts.tseries.getfield ts.tseries.repeat ts.tseries.tolist

ts.tseries.compress ts.tseries.imag ts.tseries.reshape ts.tseries.tostring

ts.tseries.conj ts.tseries.item ts.tseries.resize ts.tseries.trace

ts.tseries.conjugate ts.tseries.itemset ts.tseries.round ts.tseries.transpose

ts.tseries.copy ts.tseries.itemsize ts.tseries.searchsorted ts.tseries.var

ts.tseries.ctypes ts.tseries.max ts.tseries.setfield ts.tseries.view

ts.tseries.cumprod ts.tseries.mean ts.tseries.setflags

```

### Other Date Functions

Variations on a theme:

```

# truncation

ts.truncdate(

start=datetime(2017, 1, 1),

finish=datetime(2017, 12, 31))

# just start date etc.

ts.truncdate(

start=datetime(2017, 1, 1))

# this was in date order but suppose it was in reverse order?

# this result will give the same answer

ts1 = ts.truncdate(

start=datetime(2017, 1, 1),

new=True)

ts.reverse()

ts1 = ts.truncdate(

start=datetime(2017, 1, 1),

new=True)

# use the date format native to the dateseries (ordinal / timestamp)

ts1 = ts.truncdate(

start=datetime(2017, 1, 1).toordinal(),

new=True)

# suppose you start with a variable that represents a date range

# date range can be either a list or tuple

ts.truncdate(

[datetime(2017, 1, 1), datetime(2017, 12, 31)])

```

## Assorted Date Functions

```

# native format

ts.daterange()

(735963, 735972)

# str format

ts.daterange('str')

('2015-12-31', '2016-01-09')

# datetime format

ts.daterange('datetime')

(datetime.datetime(2015, 12, 31, 0, 0), datetime.datetime(2016, 1, 9, 0, 0))

# native format

ts.start_date(); ts.end_date()

735963 735972

# str format

ts.start_date('str'); ts.end_date('str')

2015-12-31 2016-01-09

# datetime format

ts.start_date('datetime'); ts.end_date('datetime')

2015-12-31 00:00:00 2016-01-09 00:00:00

```

Sometimes it is helpful to find a particular row based on the date. Also, that date might not be in

the dateseries, and so, the closest date will suffice.

We will create a sample timeseries to illustrate.

```

ts = Timeseries()

ts.dseries = []

ts.tseries = []

start_date = datetime(2015, 12, 31)

for i in range(40):

date = start_date + timedelta(days=i)

if date.weekday() not in [5, 6]: # skipping weekends

ts.dseries.append(date.toordinal())

ts.tseries.append(i)

ts.make_arrays()

# row_no, date

(0, '2015-12-31')

(1, '2016-01-01')

(2, '2016-01-04')

(3, '2016-01-05')

(4, '2016-01-06')

(5, '2016-01-07')

(6, '2016-01-08')

(7, '2016-01-11')

(8, '2016-01-12')

(9, '2016-01-13')

(10, '2016-01-14')

(11, '2016-01-15')

(12, '2016-01-18')

(13, '2016-01-19')

(14, '2016-01-20')

(15, '2016-01-21')

(16, '2016-01-22')

(17, '2016-01-25')

(18, '2016-01-26')

(19, '2016-01-27')

(20, '2016-01-28')

(21, '2016-01-29')

(22, '2016-02-01')

(23, '2016-02-02')

(24, '2016-02-03')

(25, '2016-02-04')

(26, '2016-02-05')

(27, '2016-02-08')

date1 = datetime(2016, 1, 7) # existing date within date series

date2 = datetime(2016, 1, 16) # date falling on a weekend

date3 = datetime(2015, 6, 16) # date prior to start of date series

date4 = datetime(2016, 3, 8) # date after to end of date series

# as datetime and in the series

existing_row = ts.row_no(rowdate=date1, closest=1)

5

existing_date = ts.closest_date(rowdate=date1, closest=1)

print(datetime.fromordinal(existing_date))

2016-01-07 00:00:00

# as datetime but date not in series

next_row = ts.row_no(rowdate=date2, closest=1)

12

next_date = ts.closest_date(rowdate=date2, closest=1)

print(datetime.fromordinal(next_date))

2016-01-18 00:00:00

prev_row = ts.row_no(rowdate=date2, closest=-1)

11

prev_date = ts.closest_date(rowdate=date2, closest=-1)

print(datetime.fromordinal(prev_date))

2016-01-15 00:00:00

# this will fail -- date is outside the date series

# as datetime but date not in series, look for earlier date

ts.closest_date(rowdate=date3, closest=-1)

# this will fail -- date is outside the date series

ts.closest_date(rowdate=date4, closest=1)

```

## Functions by Category

### Output

#### ts.as_dict()

Returns the time series as a dict with the date as the key and without

the header information.

#### ts.as_json(indent=2)

This function returns the timeseries in JSON format and includes the

header information.

#### ts.as_list()

Returns the timeseries as a list.

#### ts.header()

This function returns a dict of the non-timeseries data.

#### ts.items(fmt=None)

This function returns the date series and the time series as if it

is in one list. The term items used to suggest the iteration of dicts

where items are the key, value combination.

if fmt == 'str':

the dates are output as strings

#### ts.months(include_partial=True)

This function provides a quick way to summarize daily (or less)

as monthly data.

It is basically a pass-through to the convert function with more

decoration of the months.

Usage:

months(include_partial=True)

returns a dict with year-month as keys

#### ts.years(include_partial=True)

This function provides a quick way to summarize daily (or less)

as yearly data.

It is basically a pass-through to the convert function with more

decoration of the years.

Usage:

years(include_partial=True)

returns a dict with year as keys

#### ts.datetime_series()

This function returns the dateseries converted to a list of

datetime objects.

#### ts.date_string_series(dt_fmt=None)

This function returns a list of the dates in the timeseries as

strings.

Usage:

self.date_string_series(dt_fmt=None)

dt_fmt is a datetime mask to alter the default formatting.

### Array Manipulation

#### ts.add(ts, match=True)

Adds two timeseries together.

if match is True:

means there should be a one to one corresponding date in each time

series. If not raise error.

else:

means that timeseries with sporadic or missing dates can be added

Note: this does not evaluate whether both timeseries have the same

number of columns. It will fail if they do not.

Returns the timeseries. Not in-place.

#### ts.clone()

This function returns a copy of the timeseries.

#### ts.combine(tss, discard=True, pad=None)

This function combines timeseries into a single array. Combining in

this case means accumulating additional columns of information.

Truncation takes place at the end of rows. So if the timeseries is

sorted from latest dates to earliest dates, the older values would be

removed.

Usage:

self.combine(tss, discard=True, pad=None)

Think of tss as the plural of timeseries.

If discard:

Will truncate all timeseries lengths down to the shortest

timeseries.

if discard is False:

An error will be raised if the all the lengths do not match

unless:

if pad is not None:

the shorter timeseries will be padded with the value pad.

Returns the new ts.

#### ts.common_length(ts1, ts2)

This static method trims the lengths of two timeseries and returns two

timeseries with the same length.

The idea is that in order to do array operations there must be a

common length for each timeseries.

Reflecting the bias for using timeseries sorted from latest info to

earlier info, truncation takes place at the end of the array. That

way older less important values are removed if necessary.

Usage:

ts1_new, ts2_new = self.common_length(ts1, ts2)

#### ts.convert(new_freq, include_partial=True, **kwargs)

This function returns the timeseries converted to another frequency,

such as daily to monthly.

Usage:

convert(new_freq, include_partial=True, **kwargs)

The only kwarg is

weekday=<some value>

This is used when converting to weekly data. The weekday number

corresponds to the the datetime.weekday() function.

#### ts.extend(ts, overlay=True)

This function combines a timeseries to another, taking into account the

possibility of overlap.

This assumes that the frequency is the same.

This function is chiefly envisioned to extend a timeseries with

additional dates.

Usage:

self.extend(ts, overlay=True)

If overlay is True then the incoming timeseries will overlay

any values that are duplicated.

#### ts.trunc(start=None, finish=None, new=False)

This function truncates in place, typically.

truncate from (start:finish)

remember start is lowest number, latest date

This truncation works on the basis of slicing, so

finish is not inclusive.

Usage:

self.trunc(start=None, finish=None, new=False)

#### ts.truncdate(start=None, finish=None, new=False)

This function truncates in place on the basis of dates.

Usage:

self.truncdate(start=None, finish=None, new=False)

start and finish are dates, input as either datetime or the actual

internal format of the **dseries** (ordinals or timestamps).

If the dates are not actually in the list, the starting date will

be the next viable date after the start date requested. If the finish

date is not available, the previous date from the finish date will be

the last.

If new is True, the timeseries will not be modified in place. Rather

a new timeseries will be returned instead.

#### ts.replace(ts, match=True)

This function replaces values where the dates match an incoming

timeseries. So if the incoming date on the timeseries matches, the

value in the current timeseries will be replaced by the incoming

timeseries.

Usage:

self.replace(ts, match=True)

If match is False, the incoming timseries may have dates not found in

the self timeseries.

Returns the modified timeseries. Not in place.

#### ts.reverse()

This function does in-place reversal of the timeseries and dateseries.

#### ts.get_diffs()

This function gets the differences between values from date to date in

the timeseries.

#### ts.get_pcdiffs()

This function gets the percent differences between values in the

timeseries.

No provision for dividing by zero here.

#### ts.set_ones(fmt=None, new=False)

This function converts an existing timeseries to ones using the same

shape as the existing timeseries.

It is used as a convenience to create an empty timeseries with a

specified date range.

if fmt use as shape

usage:

set_ones(self, fmt=None, new=False)

#### ts.set_zeros(fmt=None, new=False)

This function converts an existing timeseries to zeros using the same

shape as the existing timeseries.

It is used as a convenience to create an empty timeseries with a

specified date range.

if fmt use as shape

usage:

set_zeros(self, fmt=None, new=False)

#### ts.sort_by_date(reverse=False, force=False)

This function converts a timeseries to either date order or reverse

date order.

Usage:

sort_by_date(self, reverse=False, force=False)

If reverse is True, then order will be newest to oldest.

If force is False, the assumption is made that comparing the first

and last date will determine the current order of the timeseries. That

would mean that unnecessary sorting can be avoided. Also, if the order

needs to be reversed, the sort is changed via the less expensive

reverse function.

If dates and values are in no particular order, with force=True, the

actual sort takes place.

This function changes the data in-place.

### Evaluation

#### ts.daterange(fmt=None)

This function returns the starting and ending dates of the timeseries.

Usage:

self.daterange()

(735963, 735972)

self.daterange('str')

('2015-12-31', '2016-01-09')

self.daterange('datetime')

(datetime(2015, 12, 31, 0, 0),

datetime.datetime(2016, 1, 9, 0, 0))

#### ts.start_date(fmt=None)

This function returns the starting date of the timeseries in its

native value, timestamp or ordinal.

If fmt is 'str' returns in string format

If fmt is 'datetime' returns in string format

#### ts.end_date(fmt=None)

This funtcion returns the ending date of the timeseries in its native

value, timestamp or ordinal.

If fmt is 'str' returns in string format

If fmt is 'datetime' returns in string format

#### ts.get_duped_dates()

This function pulls dates that are duplicated. This is to be used to

locate timeseries that are faulty.

Usage:

get_duped_dates()

returns [[odate1, count], [odate2, count]]

#### ts.series_direction()

if a lower row is a lower date, then 1 for ascending

if a lower row is a higher date then -1 for descending

#### ts.get_date_series_type()

This function returns the date series type associated with the

timeseries. The choices are TS_ORDINAL or TS_TIMESTAMP.

#### ts.if_dseries_match(ts)

This function returns True if the date series are the same.

#### ts.if_tseries_match(ts)

This function returns True if the time series are the same.

### Utilities

#### ts.date_native(date)

This awkwardly named function returns a date in the native format of

of the timeseries, namely ordinal or timestamp.

#### ts.row_no(rowdate, closest=0, no_error=False)

Shows the row in the timeseries

Usage:

ts.row(rowdate=<datetime>)

ts.row(rowdate=<date as either ordinal or timestamp>)

Returns an error if the date is not found in the index

if closest is invoked:

closest = 1

find the closest date after the rowdate

closest = -1

find the closest date before the rowdate

If no_error

returns -1 instead of raising an error if the date was

outside of the timeseries.

#### ts.get_datetime(date)

This function returns a date as a datetime object.

This takes into account the type of date stored in **dseries**.

Usage:

self.get_datetime(date)

#### ts.lengths()

This function returns the lengths of both the date series and time

series. Both numbers are included in case a mismatch has occurred.

#### ts.shape()

This function return the shape of the timeseries. This is a shortcut

to putting in ts.tseries.shape.

#### ts.fmt_date(numericdate, dt_type, dt_fmt=None)

This static method accepts a date and converts it to

the format used in the timeseries.

#### ts.make_arrays()

Convert the date and time series lists (if so) to numpy arrays

#### ts.get_fromDB(**kwargs)

This is just a stub to suggest a viable name for getting data from a

database.

#### ts.save_toDB(**kwargs):

This is just a stub to suggest a viable name for saving data to a

database.

## Project details

## Release history Release notifications

## Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Filename, size & hash SHA256 hash help | File type | Python version | Upload date |
---|---|---|---|

thymus-timeseries-0.1.1.tar.gz (62.4 kB) Copy SHA256 hash SHA256 | Source | None |