How well does Elementizing work?

[DLuckham]

Elementizing is used to add semantics tags to news reports for input to algorithmic trading programs. Why? Because "News moves Markets" - not just stock prices! Dow Jones is doing it, Thompson is doing it. And they are charging for it - a lot, I'm told.
See Ian Koenig's presentation at the Third EPTS Workshop, Session VII.
http://www.complexevents.com/?page_id=269
Now here's my question: How well does Elementizing work? Have there been any tests, comparative experiments?

[garethsmith]

As I see it an elementized feed delivers two advantages:

- The data is pre-parsed and delivered with discrete information, so the CEP applications can exploit the given values directly without having to infer from unstructured text, reducing the load and latency at the client end.

- The information delivered in an elementized feed is often sourced from the same providers who deliver the same information in a more conventional unstructured form. The vendors have a small time window (under embargo) to produce the elementized set thus ensuring that the delivery of both streams are synchronised (if not biased towards the elementized service, in latency terms).

This obviously works well with data that can be discretely defined (e.g. Indices, profit results, etc); and, in our experience, offers a low-latency means of integrating news data into CEP applications. However, even if the data is not so easily encapsulated the vendor pre-formatting (under embargo) removes additional computation (and thus latency) for the clients.

Gareth

--
Dr Gareth Smith
Principal Architect
Progress Software - Apama

[DLuckham]

Have there been any studies of the auto-categorization results versus human categorization results?
If humans could work fast enough, would the elementized data lead to any better trading results?
Or is it that the trading algorithms can't take advantage of "human subtleties"?

Why do I ask? Because I'm trying to make a case that the kind of auto-categorization that is used in Finance could be applied to data in other areas (using different ontologies), e.g., to determine relevance of a news story to some Intelligence goal.

Am I correct to read this:

However, even if the data is not so easily encapsulated the vendor pre-formatting (under embargo) removes additional computation (and thus latency) for the clients.

as negative towards expanding the areas of application?

[Tim Bass]

Hi David,

I mention one of the major cybersecurity risks related to "news-based trading" in this post:

http://thecepblog.com/2008/01/08/cybera ... l-markets/

Yours faithfully, Tim

[DLuckham]

Tim,
I think this is a very good point indeed. It obviously raises the question as to whether Elementizing technology is capable of classifying a news item as "Rumor, Malicious", or "Rumor, False".

Wall street News elementizing , I'm told, uses four categories of concepts:

Aboutness, Genre, Sentiment and Facts.

An example of a Genre would be Rumor. And an example of a sentiment would be Malicious.
So in theory it would be possible for an automated elementizing program to classify a news item as "Rumor, Malicious".

But would that classification make any difference to the influence of that news item on the markets?
That is, if humans hear a rumor and know it to be malicious, what do they do anyway? Act as if they believe it? Act as if the opposite is true? Wait and see?

And then of course your second point would be:
What should the trading algorithm that's using the elementized news item do?
The elementizing might work, but the trading algorithm might not.

Ah yes!! Now for the day when trading is completely automated! No humans at all! Now what does that algorithm do?

BTW, I dont think this is an example of "cyber attack". Its an example of plain old rumor mongering, isn't it?

[Tim Bass]

BTW, I dont think this is an example of "cyber attack". Its an example of plain old rumor mongering, isn't it?

Hi David,

The majority of people do not realize that the term "cyberattack" is not only used for attacks against physical or tangible objects in cyberspace; it may also used in the context of attacks against intangible objects, for example perception, knowledge, understanding.

This concept falls, generally, in the category of "information warfare" and is not limited to military operations.

In fact, there are many who believe, including myself, that cyberattacks against perception, knowledge, understanding and other intangible objects pose a much greater threat than attacks directed toward physical objects.

In addition, because attacks against tangible objects are easier to quantify and defend against, it is only natural for people to think about cyberattacks in the context of physical objects and not intangible objects, like perception or situational knowledge.

I plan to elaborate on this topic in more detail in my blog this year, as time permits.

Yours faithfully, Tim

[DLuckham]

Yes Tim, I see.
But you didn't answer the questions about what to do about it:

What should the trading algorithm that's using the elementized news item do?
The elementizing might work, but the trading algorithm might not.
That is, if humans hear a rumor and know it to be malicious, what do they do anyway?
Act as if they believe it?
Act as if the opposite is true?
Wait and see?
Ah yes!! Now for the day when trading is completely automated! No humans at all! Now what does that algorithm do?

[Tim Bass]

Hey David!

Well, I don't have the answers, that's why I didn't go deeper but I'll keep the momentum going....

Yes Tim, I see.
But you didn't answer the questions about what to do about it:

What should the trading algorithm that's using the elementized news item do?

It should seek optimal times to entry and exit the market based on an overall strategy or goal.

The elementizing might work, but the trading algorithm might not.
That is, if humans hear a rumor and know it to be malicious, what do they do anyway?
Act as if they believe it?
Act as if the opposite is true?
Wait and see?

I don't think subjective events (rumor, analysis, commentary, editorial) should be used for trading (in the foreseeable future). I do think that established indicators can be used, where hard numbers are released against projections. BTW, this is what is demonstrated in the Apama screen shot (established indexes and the variance from concensus estimates.

Ah yes!! Now for the day when trading is completely automated! No humans at all! Now what does that algorithm do?

Trading will never be completely automated. There will always be some trading (no idea how much) that will be performed by humans.

Cheers!

Yours faithfully, Tim

[DLuckham]

Tim,

There are various variants of "Elementizing" in different problem domains. The Wall Street version (which coined that term) does in fact categorize news items as possibly having the genre "rumor". (See Ian Koenig's presentation in the Sept 2007 Event Processing Workshop. ) Now, what weight trading algorithms give to "rumors" and other genres would be interesting to know. It could depend upon context - e.g., how many other similar rumors have happened in the last hour. I'm hoping we might see some answers on this forum!

That's sort of what I hoped to get discussed when I asked "how well does elementizing work".
Another question that bothers me: "is Elementizing different from Keyword search with an Onotolgy?"

[hgilde]

I don't recall hearing about any rumor in this case. The analyst gave a convincing argument about how things could turn out for E-Trade. Which is what analysts do every day. And if this analyst is trying to harm E-Trade, then yes this kind of thing has been happening since the invention of the short.

So that analyst report would show a negative sentiment and some information about what topic the analyst discusses. If you're worried about E-Trade's liquidity, you might have an algorithm waiting for a report on E-Trade's credit or holdings, having a negative sentiment. Or part of your risk limiting plan might be to watch for a negative price trend after seeing any analyst report with negative sentient.

Of course, these algorithms could be either computerized or just a trader watching the wires.

Now think about what happens when everyone has an algorithm like this. They can all trigger at the same time. Add on top of that, add traders who just follow market momentum.

Computers make this process faster, but yes, this topic has had a lot of study.