Kumesh,

Awesome. Thanks for posting. One question: Have you done research to find whether or not number of story mentions (+ 20 millions) affects the returns generated by the algorithm? Or a broader question, what impact that has on the algorithm as a whole?

Seong

Thank you for the question, Seong. We categorize a story as a company and event pair that last for a 2-week period on average before people move on to something else. So Starbucks (company) and Product Recall (event) = 1 story. A story can have multiple articles associated with it. So in this event-driven example 'Starbucks' and 'Company Earnings' is a story and the Story Volume itself can be 100 (# of articles / mentions) after x days. Since we used Story Sentiment, we're aggregating the sentiment from the 100 articles/mentions on Starbucks and Company Earnings so the sentiment is much more stable.

To answer the question, yes the number of articles mentioned in a story can affect the returns because 1) more articles = more chances of a condition being matched (ex. source rank) 2) more articles in a story = more stable and reliably sentiment score on the story. A story can have 1 article which starts off negative but as more articles come in, their views could change and the sentiment could trend into the positive.

Thanks for the prompt reply Kumesh, I have a few more if you don't mind off of the following statement:

"A story can have 1 article which starts off negative but as more articles come in, their views could change and the sentiment could trend into the positive."

In what timescale is this kind of change observable? E.g. do you see this at the millisecond level or the minutely level?

Thanks

The change in story sentiment can take place in minutes to hours depending on how popular the event is. For example, lets say as soon as Apple released a new product, some influential blogger found a fault and mentioned in an article that the product is terrible - giving the story a negative sentiment. However, other people that received the product did not find any faults, thus reporting positive things about the product. This changes and stabilized the story sentiment from negative to positive within minutes or hours based on the popularity of the product release event.

Here is a backtest using the news and blogs data on Financial Rating events which returned 322%. Seems Financial Ratings is an excellent predictor for stock price movements.

can articles be processed in real-time if so how do we code that in? Is there a different CSV I have to use?

Hi Josh,

Each article in our dataset is timestamp to the second but in this Quantopian backtest, we round it to the nearest minute. When conducting the backtest, there is a 2 minute delay from when the article gets recognized by our trading condition to it actually executing the trade. Unless you are trading high frequency, then the 2 minute trade execution time shouldn't be a problem.

Best,
Kumesh

Cloned the top algorithm by Kumesh to run it. But, it always seems to time out with this below message:
"TimeoutException: Fetcher data not processed in 360 seconds There was a runtime error on line 127."
Can anyone help?

Is there any news on when data like this can be used in live trading?

I'm also interested when the data set will be available for backtesting and live trading. Also, the CSV provided in many of the examples is the full Accern sentiment data. It looks like only the AlphaOne stream (daily data and limited fields) will be available for live trading within Quantopian.

Hi folks,

We're actively working on making this data available for live trading. We'll let you know once we have news to share.

In backtesting, a fetcher file has 6 minutes (360 seconds) to fetch from the server and load into the algo. It sounds like you're running into this time limit. Perhaps your internet connection is slow? Or you can locally break up the file into smaller chunks for faster uploading.

Alisa

Hello,

I was looking up the premium data Accern can supply for 50$/month. But it only return you two metrics? For example the source rank is not given. I am wrong here?

Regards

Emiel

The data that is now built into Quantopian provides impact_score and sentiment, rolled up at a daily frequency for US Equities.

Hi Emiel,

Yes, the Alpha One dataset contains only two, most commonly used metrics (impact_score and sentiment). It's a daily delivered dataset for retail customers.
Source_rank, along with several other metrics like story_saturation, social_shares etc., is part of our real-time streaming product (Alpha Stream). Alpha Stream is focused towards bigger trading teams. We don't have the support for Alpha Stream on Quantopian yet, and it can only be accessed using our API.

Regards,
Anshul
Co-founder, Accern

The first algo appears to be 187% in the chart. Adjusting for the ~$2M negative cash, the return is 68%, see PvR.
Leverage hits 3.25.

PvR beats the benchmark so don't let this sour you on the sentiment data and important work they are doing. Just -- eyes open.

Tried cloning this, but the Dropbox files posted above (backtest results and data files required for backtest) are missing.
Is there a newer version of the Accern algorithms posted, with current Dropbox files?

Also, if someone has run these algorithms over the last year, pl point me to the results - my backtests stop in May 2016 as i don't have the paid subscription.