crowdsource earnings predictions and the quantopian research platform
DESCRIPTION
In this presentation, we will show you examples on how to incorporate multiple sources of earning predictions into your algorithms and why the sources matter. You will also get a sneak peek of our new beta research environment - where you can use IPython notebooks to analyze curated datasets, algorithms, and backtest results.TRANSCRIPT
Estimize and QuantopianCrowdsource Earnings Predictions and the
Research Platform
QuantopianSeong Lee (about)
What are we talking about?
● EPS, Wall Street Consensus, Earnings Surprise
● Estimize, the crowdsourced financials aggregator
● Cover Estimize’s whitepaper and replicate the results in the Quantopian Research Platform
● Develop a simple trading strategy
By the end of this talk you’ll have
● Tasted the potential for crowdsourced financial data
● A sense of how the Quantopian Research Platform works and how to run parameter optimization
What you need to know
● EPS: Earnings Per Share o (Net Income - Dividends)/(Shares Outstanding)o An important factor in predicting a company’s
performanceo I.E. Did it increase from last quarter?
What you need to know
● Wall Street Consensuso Aggregated consensus of Wall Street Analysts
(mostly sell-side)o Sell-side: Investment Banks, FAs, Brokerso Buy-side: Hedge funds, mutual funds, pension
funds, VCs, Prop firms, PEso IBES (Institutional Brokers’ Estimate System)
What you need to know
● Earnings Surpriseo When announcements do better than or worse than
earnings estimateso Look at Apple Q2:
Actual: 1.66 Consensus: 1.46 Surprise: 13.7%
Summary
● EPSo Earnings Per Share
● Wall Street Consensuso Average of all Wall Street Estimates
● Earnings Surpriseo When earnings announcements differ from
expectations
Who is Estimize?
● Estimize is a financial estimates aggregatoro “Independent, buy-side, and sell-side analysts as
well as private investors and students”o EPS and Revenue estimates side-by-side with Wall
Street Consensus estimates
Revisiting Apple Q2, 2014
Crowdsourced Estimates
● What makes them different?o Diversity of contributors
Only 7% of total participants are sell-side analysts
Looking at the whitepaper claims
● Released September 24, 2013● Claim #1: “More accurate 65% of the time
when there are 20 or more contributors”● Claim #2: “Average absolute error of the
Estimize consensus is smaller than the Wall Street Consensus by 12 bp when contributors are greater than 20”
The Toolbox
Claim #1: “65% more accurate”
● Did Estimize land higher than Wall Street when it was a positive surprise?
● Did Estimize land lower than Wall Street when it was a negative surprise?
Revisiting Apple Q2, 2014
Implementation
● Compare the number of times that Estimize correctly guessed direction
● Data was preprocessed so feel free to reach out if you want steps
Wrangling our DataFrame
● We have 13,000 rows, each with this data
● So for each row, see if estimize predicted direction correctly
Loading our Data
Prediction Direction
Plot the results
What about the number of participants?● Each announcement can have as little as 1
participants or more than 167
● Apple Q2, 2014:
Plotting against N participants
Plotting against N participants
Summary
● Positive correlation between number of participants and the accuracy rate of Estimize versus the Wall Street Consensus
● Accuracy > 65% when N > 20
Claim #2: Lower absolute error
● We’re going to look at the relative delta of each estimate instead
● Steps:1. Wrangle our DataFrame/spreadsheet (add a
column)2. Plot the results against N participants
Wrangling our data
Graphing the score: Code
Graphing the score: Plot
Conclusions
● Claim #1: Positive relationship between accuracy and number of participantso Matches up
● Claim #2: Lower relative error as number of participants increaseso Matches up
● So how do we use this data?
Implementing a trading strategy
● Goals:o Write a simple algorithm to backtest our strategyo Compare the Wall Street estimates versus Estimize
estimates in generating alphao Get a sneak-peak into the Quantopian Research
Platform
The strategy
● PEAD - Post Earnings Announcement Drifto “The tendency for a stock’s cumulative abnormal
returns to drift in the direction of an earnings surprise”
● Logic:o If there is a positive surprise (actual > estimate)
Buy and hold for 1 day o If there is a negative surprise (actual < estimate)
Sell and hold for 1 day
Implementation
● Same format as Quantopian IDE:
Steps
1. Create a universe of stocks from our dataa. Only where N >= 20
2. Setup our `initialize` and `handle_data` methods
3. Run the algorithm4. Optimize our parameters and choose the
best one
Creating a universe
Initialize
Handle_data: Main Logic
Run the algorithm
Optimize parameters: Brute Force
● Our parameters: ● Run a for loop over these parameters:
o ● Redefined initialize with new params:
Results
Results
● The strategy that held a position for 3 days using the Estimize estimates had cumulative returns of 4.97% from 10/12 - 1/14
● Try:o Long onlyo Short onlyo Longer holding periods
Final thoughts and Summary
● There are more efficient ways to optimizeo Gradient descent, Walk forward optimization,
Genetic algorithms● Easier plotting tools exist (Seaborn)● Crowdsourced estimate data can be
interesting new sources of alpha● Parameter optimization/research is possible
in the Quantopian Research Platform
Questions and Notes● Email us at [email protected]
o Ask us about the iPython notebook these slides were based off!
● Visit us at Quantopiano www.quantopian.com
● Estimize Whitepaper: http://com.estimize.public.s3.amazonaws.com/papers/Estimize%20Whitepaper%20Executive%20Summary.pdf
● Deutsche Bank Paper: http://blog.estimize.com/post/80676086439/deutsche-bank-quant-research-estimize-more-timely-and