Back to Community
Avoiding the Twitter Leak with Accern Sentiment Indicators

For those of you who haven’t followed the news – Twitter earnings report was mistakenly made public on an hour before the market had closed. The earnings report was only publicly accessible for 45 seconds when scraper bots captured the information and broadcast it on around 3:07 PM. Twitter earnings fall short from expectation. This caused a panic among shareholders which led to a 20% drop in Twitter’s stock price - shaving $5 billion right off Twitter’s market cap.

We ran a backtest on over 100,000 articles related to Twitter over a span of 6 months from November 1st, 2014 to April 29th, 2015. Each article has around 25+ metrics, however, we only used the following 3 metrics in our backtest.

  • Article Sentiment (-1 to 1): This metric calculated the sentiment
    score of an article which was relevant to a company. This can be used
    as a directional trigger.
  • Event Impact Score on Entity (1 to 100): This metric calculated if
    the article would have a greater-than-1% impact on the stock on the
    same day.
  • Overall Source Rank (1 to 10): This metric calculated the timeliness
    and reposting of a source information. This can be used as a
    trust/credibility or viral factor.

Check out the results in the backtest. Over a course of 6 months, our algorithm returned 71.6% verse Twitter's benchmark return of -5.1%.

Read full story here:

Credits on building the algo: Seong Lee and Derek Tishler

Clone Algorithm
Backtest from to with initial capital
Total Returns
Max Drawdown
Benchmark Returns
Returns 1 Month 3 Month 6 Month 12 Month
Alpha 1 Month 3 Month 6 Month 12 Month
Beta 1 Month 3 Month 6 Month 12 Month
Sharpe 1 Month 3 Month 6 Month 12 Month
Sortino 1 Month 3 Month 6 Month 12 Month
Volatility 1 Month 3 Month 6 Month 12 Month
Max Drawdown 1 Month 3 Month 6 Month 12 Month
# Backtest ID: 55429317ea02b90d247f7b0a
There was a runtime error.
3 responses

I can't look into this now, but this looks suspicious; are you sure you've correctly handled the times of the data pulled in with fetch_csv? I notice the timestamps are in US/Eastern, but you haven't set the timezone to US/Eastern, so possibly the algo has five hours of lookahead bias?

Again, not 100% sure, just asking.

Timestamp is our dataset are in UTC.