HFT Myths

[hft]

Here's my question:
At what point is the trading technique classified as algorithmic vs. high frequency ?
In other words, what is the average time-in-trade where it is no longer characterized as high frequency.
100 milliseconds ?
1 second ?
5 seconds ?

Also, I've read the articles where HFT profitability has been on the decline for many firms. I wonder if they are looking to lengthen their time horizon, use different algos and move away from the millisecond trading ?

That's pure semantics, and in-trade time isn't as important as response time. If I can average a few tenths of a tick edge by hitting prices that are then gone after a few microseconds, I can accumulate probabilistic edge throughout the day even if my in-trade times account to seconds and minutes as opposed to milliseconds.

A more appropriate question might be how latency affects your profitability. There are some "low-latency" trading strategies that could still be profitable if they tacked on a few hundred microseconds of latency. Others couldn't afford tens of micros, some even less. There are none that would qualify as high-frequency/low-latency that could afford a millisecond+ of latency. A rule of thumb is that if you need to be colocated to make money that you are transacting in low-latency trading.

 

[earlyexit]

for example, the 2010 flash crash was when hft's started losing millions within a few minutes and then all shut off, no liquidity = no market = crash.

The vast majority of the HFT firms that shut down that day did it because they didn't trust the data they were receiving. Of the ones I know that kept going that day made a killing.

 

[syswizard]

Here's my question:
At what point is the trading technique classified as algorithmic vs. high frequency ?
In other words, what is the average time-in-trade where it is no longer characterized as high frequency.
100 milliseconds ?
1 second ?
5 seconds ?

Also, I've read the articles where HFT profitability has been on the decline for many firms. I wonder if they are looking to lengthen their time horizon, use different algos and move away from the millisecond trading ?

I've asked this question because I had a trading strategy that showed a nice equity curve with an hourly bar time-frame. However, commissions offset the profits, so the curve went straight down after commissions. But commisions have declined since that time, so I am thinking of dusting it off and trying it on something like 10 tick bars.
I imagine it would fire off at least 100 trades per day. So the thinking is: if I get a prop firm with low commish maybe this will be profitable now.

 

[maxima120]

Of the ones I know that kept going that day made a killing.

OR the ones had no proper programming made and had no idea how to deal with the situation )))))))))))))))))))

типа Эла Бунская которой 18 лет и она на вершине мира yahooooooooooooooo!!

 

[Lights]

The vast majority of the HFT firms that shut down that day did it because they didn't trust the data they were receiving. Of the ones I know that kept going that day made a killing.

my point is that hft's, etc don't consider themselves "waiters" and never will. they don't think they are obligated to serve a market at all times ie catch falling knives since we all know that their primary goal is to make money, and making markets/adding liquidiy is secondary. we will continue to get less efficient markets, and more volatility which is ironic in the sense that regulators still assume technology is what creates more efficient markets and better executions.

once the markets finally collapse in the future (and it will some day who knows when) the government will usurp all financial technology and take over all market making activities

 

[hft]

I've asked this question because I had a trading strategy that showed a nice equity curve with an hourly bar time-frame. However, commissions offset the profits, so the curve went straight down after commissions. But commisions have declined since that time, so I am thinking of dusting it off and trying it on something like 10 tick bars.
I imagine it would fire off at least 100 trades per day. So the thinking is: if I get a prop firm with low commish maybe this will be profitable now.

HFT firms don't really filter out strategies with < X trades per day. It will matter more what your projected Sharpe is and the overall scope of your strategy. A major issue is that as you extend holding times, you start exceeding the bounds of expertise of the firm and it's no longer a good match for either party.

But if your system is making 100ish trades per day and is fully automated, it sounds within the scope of what most HFT firms are willing to consider, at least from a very high level.

 

[syswizard]

A major issue is that as you extend holding times, you start exceeding the bounds of expertise of the firm and it's no longer a good match for either party.

I get the importance of the Sharpe measure, but I don't really understand the above. At what point does the hold time exceed the expertise of the firm ?

 

[hft]

I get the importance of the Sharpe measure, but I don't really understand the above. At what point does the hold time exceed the expertise of the firm ?

I mean if you're going to put on a position due to some fundamental or macroeconomic factors that you will hold for days/weeks/etc, the firm's management wouldn't have the expertise or desire to judge the validity of that type of strategy. Don't think that relates to what you're doing to.

Just trying to say that HFT firms wouldn't mind you doing longer-term stuff in general, as long as it's built around alphas/infra that we have edge in.

 

[supertom]

Simple example:
EdgeRequired = 0.1 ticks
MktImbalance = BidSize/(BidSize+AskSize)
TheoPrice = BidPrice + MktImbalance * (BidPrice + AskPrice)/2

MyBid = TheoPrice - EdgeRequired, rounded down to nearest tick
MyAsk = TheoPrice + Edge Required, rounded up to nearest tick

In practice, theoPrice contains signals from both the traded market (trades occurring in the market, etc) and other markets (SPY, etc).

This is an example of a signal based on trades:
http://en.wikipedia.org/wiki/VPIN

I think that there is a typo in your formula. Do you mean the following?

TheoPrice = BidPrice + MktImbalance * (AskPrice- BidPrice )/2

which is TheoPrice = (BidPrice *AskVolume + AskPrice*BidVolume)/(AskVolume + BidVolume)

Otherwise, the theoretical price is way too high.

I have questions on VPIN:

1. How do you choose the size of the volume bucket, V?
2. How do you deal with short-term market swing, which may result in lots of inventory for MM? Is VPIN able to deal with sudden market swing? Please elaborate from your own experience.
3. There is another paper which shows different view on VPIN.
http://www.sciencedirect.com/science/article/pii/S1386418113000189

 

[hft]

I think that there is a typo in your formula. Do you mean the following?

TheoPrice = BidPrice + MktImbalance * (AskPrice- BidPrice )/2

which is TheoPrice = (BidPrice *AskVolume + AskPrice*BidVolume)/(AskVolume + BidVolume)

Otherwise, the theoretical price is way too high.

I have questions on VPIN:

1. How do you choose the size of the volume bucket, V?
2. How do you deal with short-term market swing, which may result in lots of inventory for MM? Is VPIN able to deal with sudden market swing? Please elaborate from your own experience.
3. There is another paper which shows different view on VPIN.
http://www.sciencedirect.com/science/article/pii/S1386418113000189

Yes, your formula is correct.

1. You test different sizes of buckets and analyze them. That VPIN wiki article is just an example of a way to use trades to bias your market. Real alphas used are more complex and handle more edge conditions.
2. That's really the bane of market-making in general, right? You try to minimize it by hedging, in order of correlation to the product that you're making a market in, but otherwise I lose money on these cases and the algo should adjust to it accordingly.
3. I don't have access to it, but the summary seems to indicate that VPIN doesn't predict near-term volatility. VPIN is just a start, you need to analyze/implement/adjust accordingly.