Predicting break-points with TwitterHow to use Twitter in your trading strategies!

SwipeStox

www.swipestox.com 2

These factors are naturally chosen by investors for early warnings of changes in context that might influence financial markets.

Generally speaking, break-points are typically associated with factors outside the markets:

• Breaking news• Changes in socio-political context

• Information is available after the occurrence of the event and markets only need a few milliseconds to react

• You cannot predict political or social unexpected events.

Limitations of this approach:

www.swipestox.com 3

The expansion of internet and social media led to a new concept of buzz.

• Social media offers huge waves of information that is uncontrolled and uncontrollable, comprehensive and real-time.

• They precede traditional organised media. • Several researches have shown the predictive ability of this vast mass of informal

information.

• Known as “early signs”.

www.swipestox.com 4

Arnaud Vincent and Margaret Armstrong conducted an experiment to demonstrate the predictive ability of Twitter information and to introduce an alternative use of Twitter to the world.

WHY TWITTER ?

• Designed and exploited to encourage the buzz between internet users.

• Easy to write and read “tweets”.

• There is a character limit for each tweet, making them more meaningful compared to other

posts of the same topic.

• Well-known and used worldwide.

• Easy to evaluate the significance of each tweet in present, since the number of retweets are

known.

www.swipestox.com 5

Limitation:

• Tons of information on Twitter are irrelevant of the financial markets.• Therefore, the experiment focuses on the concept of “Change of context”, which is a

measure of the volatility of Twitter along the lines of the volatility of market index VIX.

www.swipestox.com 6

• To prove the existence of the link between twitter information and break-points, the experiment counts the number of words that appeared “Twitscoop” for the first time.

• Twitscoop watches the activity of Twitter of all time.

• 2 to 10 alerts per day.

• At first: the results seemed completely different from breakpoint analysis in trading strategies, where there are periods of weeks or months between alerts.

#swipestox

www.swipestox.com 7

Testing the Twitter alerts experimentally

• As to test if the triggers could be useful to traders, they developed a modest trading strategy using a genetic algorithm.

• Genetic Algorithms can adapt themselves to gradual changes but are sensitive to sudden changes and they function by

generating random changes within them. Therefore, they allow us to carry large number of runs simultaneously and hence,

compare the results.

• By working with high frequency foreign exchange data, it is potentially easier to find out whether stopping trading after each

Twitter alert has a positive effect on performance even with data over a relatively short period. As a result, the number of

runs and the high frequency, which is one action per two minutes compensates for the short testing period of five months.

www.swipestox.com 8

The performance of a basic genetic algorithm compared to the performance of a hybrid algorithm that takes into consideration the Twitter alert.

Hybrid Algorithm

Whenever an alert occurs, the genetic algorithm stops for trading for a short time and starts again a little later after going through a new learning phase.

Forex Data

Forex data seemed the best suited for our tests because the market is highly liquid and because it is easy to obtain reliable historic data. We set up a very basic algorithm in which the $/€ trader which makes a decision every minute as to whether to convert his/her wealth into the other currency or to keep it in the same currency. To make this decision the trader has data for the previous 120 time steps on the four main currencies (CHF, GBP, JPY, USD).

#swipestox 9www.swipestox.com

Procedure

• The experiment was conducted over a 5-month period

• 100,000 observations.

• Chose randomly set of individuals with random genes.• Carried out 450 runs, based on exactly the same data and

under the same conditions.

• Results were averaged for sets of 10 runs.

Results

• Without using the Twitter alerts and overall of 450 runs, the genetic algorithm made a minor profit of 0.56%.

www.swipestox.com 10

Hybrid Algorithm

It was created to follow the objective of using the volatility measured on Twitter and the resulting alerts to improve the trading performance.

• Stop sending orders for a certain time (in this experiment 10 minutes) after an alert. • The appearance of two new words in Twitscoop constitutes an alert and should be treated as a potential breakpoint

in the series.

Result

• The average gain rose from 0.56% to 1.27%.

www.swipestox.com 11

Twitter WaveThe final test of the experiment was to recalculate the results of the hybrid algorithm for time differences ranging from -10 minutes up to +10 minutes and carried out again 450 runs for each possible time difference.

ResultThe graph shows the percentage gain/loss on average over 450 runs, if the Twitter alert had been made several minutes earlier or later. The maximum gain occurs if the information had been known 4 minutes earlier.

www.swipestox.com 12

• The preliminary outcomes from the experiment are promising.

• Open up new perspectives for identifying break-points in the performance of high frequency trading strategies by using information available to the public from Twitter.

• The results of the experiment prove the existence of a correlation between the buzz seen on Twitter and variation in the exchange rates between 4 to 6 minutes prior to the breakpoint occurrence; Twitter Wave.

“The launch of a new approach of using Twitter, which allows the traders to “surf” over the break-points and increase the profits”

www.swipestox.com 13

Thank you !

Follow us:

#swipestox