The previous post described how I went about cleaning up some yfd data using Python and numpy. I have no doubt it can be done in fewer lines of code, but I think the post described how useful it can be to manipulate arrays rather than looping through everything. With the data cleaned up, I hoped to visualize my newborn son's sleep schedule. I recently received an example that does the same thing as my python code, but in 3 lines! It uses R, ggplot2, and plyr. A few more lines can generate pretty plots like this (box plot of sleep length in hrs vs. start time):

As noted in a previous post, your.flowingdata.com (yfd) is a handy way to collect personal data. I've been collecting sleep, diaper, etc. data on my newborn son. Although yfd now allows users to calculate durations between specified events, the visualization of the information isn't quite to my liking and it's clear that errors in the data can make for some odd durations (e.g., my son slept for two days!). Numpy to the rescue!

For those of you who don't know, numpy is python's powerful array package. Rather than loop myself to death, I thought it made more sense to use of numpy's powerful slicing and masking features to clean up the data. These features make it easy to find data entry errors.

I use the Enthought python distribution for convenience sake (and because I can't resist all those libraries -- most of which I'll never use).   Below you'll find some screenshots that step through my little script. Refer to the complete code here. (Well, it's just a start really). The code is probably a bit verbose for what it does, but we all start somewhere.

The first step is getting the data into an array you can manipulate. For your reference, your.flowingdata yields data that looks like this:

array([('gmorning', '', '', '2009-10-24 23:45:36'),('gmorning', '', '', '2009-09-30 18:15:04'), ('gnight', '', '', '2009-09-23 21:00:03'), ('gmorning', '', '', '2009-09-23 19:15:03')])

I won't step through all the code here since it's available above, but you get the idea. One thing to watch out for: datetimes. I spent a lot of time trying to figure out the best way to handle the timestamps included with the yfd event data. There are ways to convert strings to ordinal numbers into datetime objects and back again, but really I wanted to manipulate the datetime objects directly to take advantage of numpy's array slicing and arithmetic. Luckily, numpy allows object types (technically, they are called 'dtypes'). This allows you to subtract one timestamp array from another to get the elapsed time without any conversions (though you'll have to convert at some point if you want to generate a human-readable string). Here's an example of the array you'll get at the end (heads -> sleep duration, start time, end time):

Just saw this...

In a previous post, I complained about the DTCC's CDS data website and the one week lifespan of the data published there. For those of you who don't know, the DTCC clears and settles a massive number of transactions every day for multiple asset classes. It's one of those financial institutions that doesn't get much press but underpins the entire capital market.

Anyway, the recent crisis motivated the DTCC to publish weekly CDS (single name, index, and tranche) exposure data. A good idea, until one realizes the data goes up in smoke when the next week's data arrives. Although DTCC recently added links to data for "a week ago", "a month ago", and "a year ago," it's still pretty inconvenient. So, if you want the data, you have to parse it yourself. I originally wanted to write a smart parser that would dynamically react to whatever format it encountered...I came to my senses and adopted a simpler approach.

The approach thus far:

• Download the raw html pages/files via "curl." Urllib2 is the preferred method to pull web pages, but I didn't have the patience to figure out how to handle redirects. Curl is a utility included with OS X that, for whatever reason, ignores redirects automatically. As such, I created a short python script to download the html for all the tables of interest weekly.
• Use BeautifulSoup to parse the html. Other libraries, such as html5lib and lxml seem to be gaining ground on BeautifulSoup, particularly as it's author wants to get out of the parsing game altogether. Nevertheless, I couldn't be bothered to figure out the unicode issues I experienced with html5lib or lxml's logic. BeautifulSoup is straightforward and "gives you unicode, dammit!" (quoting the author).
• Use numpy for easier data manipulation. Since my html, css, DOM, etc. knowledge is basic, I thought it might be better to use numpy to manipulate the table data rather than rely solely on the parser. This meant vectorizing the html data into a 1D array, cleaning it up, and generally preparing it for future reshaping. Numpy, how did I ever live without you?

This would've been much easier if all the tables were exactly the same format. Unfortunately, that's never the case. An extra cell here or there, or weird characters, can throw things off. This isn't an issue if you are parsing individual pieces of data or a single table. But what if you need to parse ten, 20, 100, etc. tables? It can get ugly fast. The DTCC data is broken into 23 pages, some of which have multiple tables. Luckily, most of my pain was self-inflicted (hey, I'm a parsing virgin). I only had to account for a few different table formats in the end.

One downside to my approach is I do not dynamically produce headers for the data I'm pulling. I plan to manually set the headers for each table (the ultimate destination for the data right now are csv files). If there's a better way, please let me know.

You can find the code here via pastebin (feedback is welcome).
You can find the DTCC tables here (if you want to view the html source).

Part 2 will cover the process of reformatting the data with numpy and perhaps feature some charts. I'm very curious to see what the numbers show!

Here are a few screenshots of a terminal session using the code so far:

I had no idea FriendFeed was driven by python!

 http://www.tornadoweb.org/

Although I'm beginning to think that it may make more sense to use something like PyTables to store time series data, it's hard to escape the ubiquity of relational databases in the enterprise. In tightly controlled corporate environments, PyTables might not even be an option. Since I'm on a database kick, I thought I might as well investigate ORMs (object relational mappers) and see whether they make sense (from an analyst perspective). SQLAlchemy (SQLA) is one of the big kahunas in the python community, though there are clearly many others (Django ORM, SqlObject, etc.).

I've come to realize SQLAlchemy doesn't promise that you'll write less code. It just promises that the additional code you write (when necessary) will be worth the additional power and flexibility it provides. SQLAlchemy allows the user to leverage the powerful idioms of the python language, provides a consistent "API" for multiple databases, and automates many database housekeeping details (e.g. transactions, joins, etc.). It also offers powerful reflection features that make accessing legacy databases simple. Furthermore, SQLAlchemy features an SQL expression language separate from the ORM so users can choose between SQL-like syntax or objects when appropriate, allowing the user to map tables to classes at will. Other ORMs bind tables and classes together tightly (a la the ActiveRecord pattern featured in Rails and other ORMs).

The documentation for SqlAlchemy is mostly good. It's good because it exists, it's maintained, and documents the complete API. The tutorials are instructive, but I felt they were a bit hard to follow since the author attempts to highlight different ways to do the same thing. This conflation of demo and tutorial makes it harder to keep track of what exactly needs to be instantiated and when. A separate interactive demo screencast + a more linear tutorial might have worked better.

Anyway, I won't cover SqlAlchemy's Expression Language here, it's available in the documentation and should make sense to those already familiar with SQL. The expression language essentially transforms SQL into method calls (e.g. "table.insert().values(values)" rather than "INSERT INTO table (fields) VALUES (values)").

So, using matplotlib's handy quote_historical_yahoo() function, we can replicate the database from a previous post using SqlAlchemy's declarative plugin. The declarative plugin allows the user to map tables to objects in a single step. The following code defines two tables, "assets" and "prices," and defines a function for pulling data for a given ticker from yahoo (adapted from the previous post on sqlite and python).

'''
SQLAlchemy ORM declarative example.
'''

from sqlalchemy.ext.declarative import declarative_base
from sqlalchemy import Table, Column, Integer, String, DECIMAL
 from sqlalchemy import MetaData, create_engine, ForeignKey
 from sqlalchemy.orm import relation, backref, sessionmaker, scoped_session
from matplotlib.finance import quotes_historical_yahoo
import datetime
import os

path = os.path.expanduser('~')  + \
     '/Dev/Data/AssetPrices/SQLite/assetprices_SA.sqlite'
 engine = create_engine('sqlite:////' + path, echo=True)
Base = declarative_base(bind=engine)

date1 = datetime.datetime(2009,1,1)
date2 = datetime.datetime.now()

#DEFINE TABLES#

class Asset(Base):
    __tablename__ = 'assets'
    
    asset_id = Column(Integer, primary_key=True)
    ticker = Column(String, unique=True)
    tag = Column(String)
    
    prices = relation('Price', order_by='Price.gregorian_day', backref='assets')
    
    def __init__(self, ticker, tag):
        self.ticker = ticker
        self.tag = tag
        
    def __repr__(self):
        return "<Asset('%s', '%s')>" % (self.ticker, self.tag)
    
class Price(Base):
    __tablename__ = 'prices'
    
    price_id = Column(Integer, primary_key=True)
    asset_id = Column(Integer, ForeignKey('assets.asset_id'))
    gregorian_day = Column(Integer)
    date_string = Column(String)
    year = Column(Integer)
    month = Column(Integer)
    day = Column(Integer)
    px_open = Column(DECIMAL)
    px_close = Column(DECIMAL)
    px_high = Column(DECIMAL)
    px_low = Column(DECIMAL)
    volume = Column(Integer)
    
    #asset = relation(Asset, backref=backref('prices',
                                            #order_by=gregorian_day))

    def __init__(self, gregorian_day, date_string, year, month, day,
                 px_open, px_close, px_high, px_low, volume):
        self.gregorian_day = gregorian_day
        self.date_string = date_string
        self.year = year
        self.month = month
        self.day = day
        self.px_open = px_open
        self.px_close = px_close
        self.px_high = px_high
        self.px_low = px_low
        self.volume = volume
        
    def __repr__(self):
        return "<Price('%s', '%s', '%s','%s','%s','%s','%s','%s','%s','%s')>" \
               % (self.gregorian_day, self.date_string, self.year, self.month,
                  self.day, self.px_open, self.px_close, self.px_high, 
                  self.px_low, self.volume)

#CREATE DB TABLES#
 
Base.metadata.create_all(engine)

#PACKAGE RAW DATA#

def package_data(db=None, ticker=None, tag='stock', start=None, end=None):
    '''
    package_data() uses quotes_historical_yahoo() to create a data set for a 
    given stock's price history. Date_string, Year, month, and day fields are 
    included for added flexibility. Returns a dictionary of tuples.
    '''
    raw_quotes = quotes_historical_yahoo(ticker, start, end) #list of tuples
    
    data = []
    for quote in raw_quotes:
        date_raw = datetime.datetime.fromordinal(int(quote[0]))
        year, month, day = date_raw.year, date_raw.month, date_raw.day
        date_string = date_raw.strftime("%Y-%m-%d")
        record = (ticker, tag, quote[0], date_string, year, month, day,
                  quote[1], quote[2], quote[3], quote[4], quote[5])
        data.append(record)    
    
    headers = ('ticker',
               'tag',
               'gregorian_day', 
               'date_string', 
               'year', 
               'month', 
               'day', 
               'px_open', 
               'px_close', 
               'px_high', 
               'px_low', 
               'volume')
    return {'data':data, 'headers':headers}
 

Executing this code essentially sets up the schema for an sqlite database and provides a package_data() function that will pull in data for a given ticker and date range. The "echo=True" parameter in  "engine = create_engine('sqlite:////' + path, echo=True)" will print out the SQL statements SQLA generates to the terminal.

Executing the code yields:

>>>
Evaluating SAscratchcode.py
2009-08-14 00:58:57,862 INFO sqlalchemy.engine.base.Engine.0x...b9b0 PRAGMA table_info("assets")
2009-08-14 00:58:57,863 INFO sqlalchemy.engine.base.Engine.0x...b9b0 ()
2009-08-14 00:58:57,863 INFO sqlalchemy.engine.base.Engine.0x...b9b0 PRAGMA table_info("prices")
2009-08-14 00:58:57,863 INFO sqlalchemy.engine.base.Engine.0x...b9b0 ()
2009-08-14 00:58:57,864 INFO sqlalchemy.engine.base.Engine.0x...b9b0
CREATE TABLE assets (
    asset_id INTEGER NOT NULL,
    ticker VARCHAR,
    tag VARCHAR,
    PRIMARY KEY (asset_id),
     UNIQUE (ticker)
)

2009-08-14 00:58:57,864 INFO sqlalchemy.engine.base.Engine.0x...b9b0 ()
2009-08-14 00:58:57,865 INFO sqlalchemy.engine.base.Engine.0x...b9b0 COMMIT
2009-08-14 00:58:57,866 INFO sqlalchemy.engine.base.Engine.0x...b9b0
CREATE TABLE prices (
    price_id INTEGER NOT NULL,
    asset_id INTEGER,
    gregorian_day INTEGER,
    date_string VARCHAR,
    year INTEGER,
    month INTEGER,
    day INTEGER,
    px_open NUMERIC(10, 2),
    px_close NUMERIC(10, 2),
    px_high NUMERIC(10, 2),
    px_low NUMERIC(10, 2),
    volume INTEGER,
    PRIMARY KEY (price_id),
     FOREIGN KEY(asset_id) REFERENCES assets (asset_id)
)

2009-08-14 00:58:57,866 INFO sqlalchemy.engine.base.Engine.0x...b9b0 ()
2009-08-14 00:58:57,868 INFO sqlalchemy.engine.base.Engine.0x...b9b0 COMMIT

This output shouldn't be too surprising. We've basically just created the tables we defined. So let's experiment interactively and create an asset object for Google.

>>> GOOG=Asset('GOOG', 'stock')
>>> GOOG
<Asset('GOOG', 'stock')>
>>> GOOG.ticker
'GOOG'

As you can see, it's possible now to call attributes of the GOOG object by name (e.g. ticker).

In our table definitions, we used SQLA's relation() function (e.g. prices = relation('Price', order_by='Price.gregorian_day', backref='assets')) to define a one-to-many relationship between an asset and its prices. SQLA uses the foreign key defined in the prices table to automatically generate the correct SQL. This is particularly interesting for sqlite users as sqlite parses foreign key statements but does not enforce them. Using this relation function, we can actually use dot notation to look at the GOOG objects prices attribute, just as if the prices are part of the object.

>>> GOOG.prices
[]

The prices are represented by an empty lists since we haven't actually written any prices into the database yet. So insert some prices.

>>> raw = package_data(ticker='GOOG', start=date1, end=date2)
>>> raw['headers']
('ticker', 'tag', 'gregorian_day', 'date_string', 'year', 'month', 'day', 'px_open', 'px_close', 'px_high', 'px_low', 'volume')

The package_data() function returns a python dictionary, {'data':[(list of tuples)], 'headers',(tuple of headers)}. So how do we assign the prices to the right ticker? The obvious way to do it would be to use a list comprehension to create a list of Price objects, and assign them to the GOOG object's "prices" attribute.

>>> GOOG.prices = [Price(record[2],record[3],record[4], record[5],record[6],record[7],record[8],record[9],record[10],record[11]) for record in raw['data']]
>>> GOOG.prices
[<Price('733409.0', '2009-01-02', '2009','1','2','308.6','321.32','321.82','305.5','3610500')>,
<Price('733412.0', '2009-01-05', '2009','1','5','321.0','328.05','331.24','315.0','4889000')>,...]

I've just listed the first two records to save some space, but you get the picture. Now, the import things to recognize here is that no SQL has been issued to the database yet. In order to reduce the back and forth between the database, SQLA uses a Session() object to keep track of what's going on. So let's setup a session and add our GOOG object to the session so SQLA can track it.

>>> Session = scoped_session(sessionmaker(engine))
>>> session = Session()
>>> session.add(GOOG)
>>> session.commit()
2009-08-14 01:10:15,424 INFO sqlalchemy.engine.base.Engine.0x...b9b0 BEGIN
2009-08-14 01:10:15,425 INFO sqlalchemy.engine.base.Engine.0x...b9b0 INSERT INTO assets (ticker, tag) VALUES (?, ?)
2009-08-14 01:10:15,425 INFO sqlalchemy.engine.base.Engine.0x...b9b0 ['GOOG', 'stock']
2009-08-14 01:10:15,471 INFO sqlalchemy.engine.base.Engine.0x...b9b0 INSERT INTO prices (asset_id, gregorian_day, date_string, year, month, day, px_open, px_close, px_high, px_low, volume) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
2009-08-14 01:10:15,471 INFO sqlalchemy.engine.base.Engine.0x...b9b0 [1, 733409.0, '2009-01-02', 2009, 1, 2, '308.6', '321.32', '321.82', '305.5', 3610500]
2009-08-14 01:10:15,472 INFO sqlalchemy.engine.base.Engine.0x...b9b0 INSERT INTO prices (asset_id, gregorian_day, date_string, year, month, day, px_open, px_close, px_high, px_low, volume) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
2009-08-14 01:10:15,472 INFO sqlalchemy.engine.base.Engine.0x...b9b0 [1, 733412.0, '2009-01-05', 2009, 1, 5, '321.0', '328.05', '331.24', '315.0', 4889000]
...(and it continues on)...

So the code above basically creates a session, adds our GOOG object to the session, and then commits all changes. The commit() method signals SQLA to issue all the necessary SQL in a single transaction to our sqlite database. Now that there are actual prices in the database, we can check out GOOG.prices:

>>> GOOG.prices[0]
2009-08-14 01:18:57,881 INFO sqlalchemy.engine.base.Engine.0x...b9b0 BEGIN
2009-08-14 01:18:57,883 INFO sqlalchemy.engine.base.Engine.0x...b9b0 SELECT assets.asset_id AS assets_asset_id, assets.ticker AS assets_ticker, assets.tag AS assets_tag
FROM assets
WHERE assets.asset_id = ?
2009-08-14 01:18:57,883 INFO sqlalchemy.engine.base.Engine.0x...b9b0 [1]
2009-08-14 01:18:57,885 INFO sqlalchemy.engine.base.Engine.0x...b9b0 SELECT prices.price_id AS prices_price_id, prices.asset_id AS prices_asset_id, prices.gregorian_day AS prices_gregorian_day, prices.date_string AS prices_date_string, prices.year AS prices_year, prices.month AS prices_month, prices.day AS prices_day, prices.px_open AS prices_px_open, prices.px_close AS prices_px_close, prices.px_high AS prices_px_high, prices.px_low AS prices_px_low, prices.volume AS prices_volume
FROM prices
WHERE ? = prices.asset_id ORDER BY prices.gregorian_day
2009-08-14 01:18:57,885 INFO sqlalchemy.engine.base.Engine.0x...b9b0 [1]
<Price('733409', '2009-01-02', '2009','1','2','308.6','321.32','321.82','305.5','3610500')>

Using normal python slicing syntax, we've just called up the first record in the prices table for Google. In this case, SQLA uses "lazy loading" to pull the appropriate Price object by issuing the SQL on demand. Users can choose to 'eager load' the data as well. Now that the corresponding Price object has been pulled we can inspect other attributes.

>>> GOOG.prices[0].px_high
Decimal("321.82")
>>> test_run = [(record.date_string, record.px_close) for record in GOOG.prices[0:10]]
>>> test_run
[(u'2009-01-02', Decimal("321.32")), (u'2009-01-05', Decimal("328.05")), (u'2009-01-06', Decimal("334.06")), (u'2009-01-07', Decimal("322.01")), (u'2009-01-08', Decimal("325.19")), (u'2009-01-09', Decimal("315.07")), (u'2009-01-12', Decimal("312.69")), (u'2009-01-13', Decimal("314.32")), (u'2009-01-14', Decimal("300.97")), (u'2009-01-15', Decimal("298.99"))]

In the example above, we call up the high price for the first record in our table. 'test_run' simply creates a list of tuples, using the date_string and px_close fields.

Anyway, there's a lot more to SQLA, this just scratches the surface. We'll see how deep the rabbit hole goes!

I recently talked to someone who was interested in integrating the different MS-Office products programmatically. The obvious solution is VBA, since it's built-in. I have no desire to learn VBA, but Python does offer the win32 COM interface. I'd almost forgotten since I've been using a Mac for a while. Anyway, I ran across this short tutorial on using COM and Python to automate the creation of powerpoint slides. I used COM with excel a while back, but it was slow (and thus turned to the very nice xlwt/xlrd combo to manipulate excel files). Nevertheless, I can see it coming in handy if you are constantly updating slides with essentially the same, but more recent, data.

UPDATE: I recently learned of reStructuredText, which many python tools use to create documentation from plain text files. There are tools such as S5, Bruce, and rst2pdf that facilitate the creation or display of presentations in different ways.

I'm wondering whether it makes more sense to store time series data in sqlite or a hierarchical database (like PyTables, which is based on the HDF5 format). Or maybe even straight-up text files!

Sqlite is nice because it runs everywhere and can connect to almost anything. Could serve as the 'Rosetta Stone' for slinging data around.

But PyTables is nice because it integrates with multidimensional numpy arrays and offers object-like convenience, meaning inter-row analysis is probably easier. Pytables is probably faster than sqlite but that's not really a big concern for me. Both hold plenty of data.

Text files, like CSV, are dead-simple and immediately accessible, but would require more logical work.

Decisions, decisions...

Of course, I forgot to attach the code to the previous post.

The previous post showed how matplotlib can pull market data from Yahoo. Using some python-fu, you can easily create CSV files for given stocks. CSV files are great as they are easy to parse and don't require a lot of overhead (in terms of setting things up, you can just open the file directly). Of course, Yahoo, Google, and others offer decent tools to view stock prices, so creating CSV files might be a bit redundant. Furthermore, if you really want to go whole hog and pull prices for thousands of assets, I imagine one would have to think carefully about an appropriate filesystem.

On the other hand, python includes a built in wrapper for sqlite, the ubiquitous file-based database that's embedded in so many mobile and desktop applications. Sqlite runs on any platform and is a completely open source, flyweight piece of software (which explains it's popularity). This post will outline how to create a simple two table sqlite database to store stock prices.

First if all, sqlite has it's own interpreter you can access from the command line, so if you're an sql cowboy, you can instantly execute sql statements from the prompt to explore a given database. Here are a few examples:

The .help command lists some of the sqlite specific commands available at the command line.

The .schema command prints the tables in the database (just as you might think).

You can also input sql statements directly.

Anyway, enough about sqlite itself, let's get to the python that built the database above. The attached script used four functions to package data for a defined set of tickers, using quotes_historical_yahoo(), and inserts the data into an sqlite database. The key is using a dictionary of tuples to pass around data.

stocks = {'data':[('GOOG', 'stock'), ('AAPL','stock')],
          'headers':('ticker', 'tag')}

In the example above, the 'data' key references a list of tuples that provide information on the ticker and type of asset in the tuple. The 'header' key references the table field names (as defined in the database schema) associated with the 'data' key. The same structure is used to write market data to the database, where the 'data' key refers to numeric data (e.g., foreign keys, date string, open, close, etc.) and the 'header' key refers to the formal fields to which each piece of data is associated.

raw_quotes = quotes_historical_yahoo(ticker, start, end) #list of tuples
data = []
for quote in raw_quotes:
    date_raw = datetime.datetime.fromordinal(int(quote[0]))
    year, month, day = date_raw.year, date_raw.month, date_raw.day
    date_string = str(year)+'-'+str(month)+'-'+str(day)
    record = (f_key, quote[0], date_string, year, month, day,
              quote[1], quote[2], quote[3], quote[4], quote[5])
    data.append(record)    
 
headers = ('asset_id', 
           'gregorian_day', 
           'date_string', 
           'year', 
           'month', 
           'day', 
           'open', 
           'close', 
           'high', 
           'low', 
           'volume')
return {'data':data, 'headers':headers}

I use a fair amount of string substitution in the code, which isn't strictly recommended, but I'm not really concerned with security. Rather than create two separate functions for adding stocks to the asset table and market data to the prices table, I used string operations to expand or contract the substitution marks to match the number of headers. I'm sure there are more elegant ways to do it, but, hey, it works.

Finally, I'll plug the Sqlite Manager add-on for Firefox. It's a basic and easy way to inspect sqlite files.