Python for Quant Finance

how to build an energy trading business from scratch

Dr. Teodora Baeva

In [1]:
% pylab inline

Populating the interactive namespace from numpy and matplotlib

What does a quant do in commodities trading?

  • model development
  • portfolio optimisation
  • trading strategies
  • pricing of complex products (storage, tolling, swing contracts)

Example: Gas storage

  • natural gas can be stored
  • used to
    • meet load variations
    • speculation (buy low, store, sell high)
  • subject to physical constraints...
    • injection and withdrawal rate depend on time and current amount of gas stored
    • minimum and maximum storage capacity can vary with time
  • ...find an optimal strategy
In [2]:
import numpy as np
import pandas as pd
from datetime import datetime

start_date = datetime(2015, 7, 1)
end_date = datetime(2016, 7, 1)
dates = pd.date_range(start_date, end_date)

max_inventory = pd.TimeSeries(3640, index=dates)
max_inventory[10:] = 3200
min_inventory = pd.TimeSeries(0, index=dates)
min_inventory[10:] = 200
with_rate = pd.Series([2, 4, 8, 15, 23, 30, 35, 38, 40, 42, 43, 43], 
                      index=[0, 1300, 1550, 2000, 2150, 2550, 2800, 3100, 3150, 3200, 3350, 3600])
inj_rate  = pd.Series([27, 26, 25, 24, 23, 22, 21, 20, 17, 11, 10, 5], 
                      index=[0, 1300, 1550, 2000, 2150, 2550, 2800, 3100, 3150, 3200, 3350, 3600])
wth_rate_constraints = pd.DataFrame({'min': pd.Series(2, index=dates),
                                     'max': pd.Series(45, index=dates),
                                    })
inj_rate_constraints = pd.DataFrame({'min': pd.Series(2, index=dates),
                                     'max': pd.Series(26, index=dates),
                                    })
In [3]:
storage_params = {'name': 'TestStorage',
                  'start_inventory': 2500, 
                  'max_inventory': max_inventory,
                  'min_inventory': min_inventory,
                  'with_rate': with_rate,
                  'inj_rate': inj_rate,
                  'inj_cost': 0,
                  'with_cost': 0,
                  'end_inventory': 2500,
                  'storage_start_date': start_date,
                  'storage_end_date': end_date,
                  'wth_rate_constraints': wth_rate_constraints,
                  'inj_rate_constraints': inj_rate_constraints,
                 }
In [4]:
import btglib.quant.models.storage as sc

storage = sc.Storage(storage_params)
In [5]:
curve = 20*pd.Series(np.sin(np.linspace(0,10,len(dates))/1.5-0.5), index=dates)+40

storage.optimize(curve, curve)
storage.plot();
In [6]:
curve = 20*pd.Series(np.sin(np.linspace(0,10,len(dates))/1.5+1), index=dates)+40

storage.optimize(curve, curve)
storage.plot();

The optimal strategy depends on the gas price curve

...which we don't know...

...so we generate scenarios

In [7]:
from btglib.quant.models.simulator import Simulator

model_parameters = {'distribution': 'LogNormal',
                    'volatility': np.array([0.1]*len(dates))[:,None], 
                    'correlation': np.eye(1), 
                    'names': ['gas'],
                   }

simulator = Simulator('5-Jun-2015', dates, curve[:,None], model_parameters)
sims = simulator.generate_paths(100)
In [8]:
plt.plot(sims.ix[0,:,:], color='k', alpha=0.1);
plt.plot(curve, color='r', linewidth=2);

Stochastic optimisation problem

  • optimal strategy under uncertainty
  • Monte-Carlo simulation with different distributions
  • large optimisation problems
  • price sensitivities

Why Python?

  • ideal for start-up
  • quick cycle from idea to prototype to tool
  • "batteries included" - scientific libraries, visualisation tools, documentation
  • python may be slow...but cython (c/c++) is not!

To trade complex products one first needs to be able to...trade!

In [9]:
trades = pd.DataFrame({'Product': ['WkEnd 04/07-05/', 'Q315', 'Aug 15'], 
                       'Market': ['Germany Baseload', 'Germany Peaks', 'Germany Baseload'], 
                       'Volume': [50, -10, 20]})
  • Building an ETRM (Energy Trading and Risk Management system)
    • parse trades
    • calculate positions and PnL per trader/book/strategy
    • produce custom reports
In [10]:
trades
Out[10]:
Market Product Volume
0 Germany Baseload WkEnd 04/07-05/ 50
1 Germany Peaks Q315 -10
2 Germany Baseload Aug 15 20
In [11]:
from btglib.utils import parse_traded_product

schedules = [parse_traded_product(p, m, datetime.now(), v) for p,m,v in trades[['Product','Market','Volume']].values]
position = pd.concat(schedules, axis=1).sum(axis=1)
position.plot();

ETRM

  • trade processing with re
  • positions generation/aggregation with pandas
  • reporting with reportlab
  • interface with Excel using pyxll

To protect downside risk...need to be able to trade options!

(and to make money too)


  • model for quick valuation of option portfolios
  • display portfolio sensitivities
  • pnl attribution
  • scenario testing

All of these require an implied volatility surface

In [12]:
from IPython.display import HTML
HTML('<iframe src=http://data.theice.com/MyAccount/Login.aspx?ReturnUrl=~\ViewData\Default.aspx width=1060 height=550></iframe>')
Out[12]:

Web-scraping with Python

  • requests for simple web pages
  • selenium when we need to drive a web browser to navigate to the data
    • portable software testing framework for web applications
    • compatible with most modern browsers
    • bindings for a number of popular programming languages, including Python (and some others like Java, C#, Groovy, Perl, PHP and Ruby)
  • BeautifulSoup to parse the data
In [13]:
from IPython.display import HTML
HTML('<iframe src=https://plot.ly/~tbaeva/98/vol-surface-25-feb-2015/ width=960 height=700></iframe>')
Out[13]:

Automation, reporting, alerting

  • python is great for automating tasks to scrape data, process and produce model results
  • ...but also to produce and send out custom reports
  • reports can be built with reportlab and saved as pdf
  • emailing is easy with smtplib

Automation, reporting, alerting

  • reading your Outlook to automatically process data (e.g. from brokers)
In [14]:
import win32com.client

outlook = win32com.client.Dispatch("Outlook.Application").GetNamespace("MAPI")
inbox = outlook.GetDefaultFolder(6)
last_message = inbox.Items.GetLast()
print last_message.sender()
print last_message.subject

Yves Hilpisch
RE: OSQF Conference in Frankfurt 05.06.2015

Deploying models to the desk

  • traders love Excel
  • python modules can be used from Excel, e.g. via pyxll (Tony Roberts)
  • easy to centralise application modules and keep them under version control
  • however...
    • requires python installation on trader's machine
    • package updates locally
    • ...and traders don't like rebooting

Deploying models to the desk

  • web interface
    • models are triggered from the browser
    • models run on a server
    • results displayed on browser
    • model runs can be logged
    • easy to embed interactive plotting!

Web Interface

  • built using flask and jinja
  • allows to have interface, documentation, wiki etc in one place
  • running via Apache

Web Interface

  • interactive plotting with bokeh

Web Interface

  • embedded pdf report created with reportlab

In conclusion

  • python is traditionally seen as a scientific language (scientific libraries)
  • due to python's versatility and rich ecosystem it can be used to build a trading business from scratch
  • full suite from data collection, to modelling, to visualisation, to automation, reporting and alerting, to great user interfaces


thanks!