@ Asim, yes, it's is absolutely important to model the divergence risk.In my experience it is the most complicated but also the most rewarding part of the pairs trading strategy. Indeed, the signal confirmation approach is the one that comes naturally. By the way, you write (paraphrased), "enter the pair if the signal confirmed with chance of convergence before close". One caveat, as I wrote, it depends on the market itself: i.e. liquidity, market participants, their IT, speed etc. Unless you have fastest IT in the market, you don't want to be trading something that is likely to converge very soon, e.g. within 1 sec. That's more pure arbitrage than stat arb. That's why it's rewarding to play stat arb strategies as they are less dependent on the speed and more on the smart application of stats.

I disagree that your model has to become more sophisticated to generate alpha. A simple robust approach will always be better. It will also help to stop you from curve fitting. You should always be thinking KISS.

The typical EOD pairs strategy is popular because it's easier to backtest. You can trade it intraday but you have to backtest it.

Remember most pairs trading strategies are looking for stationarity either directly or indirectly. When going to an intraday strategy from an EOD one, results with you targeting a different frequency. You have to ensure that there exists opportunities at this new frequency.

Also you might identify a MR signal intraday but once you take into consideration your costs and slippage there might not be enough alpha remaining to actually trade it.

My $.02

I have been trading pairs since 90's. It's still a great trade provided you do your statistical and quantitative analysis correctly. But as I said, transaction costs are critical. You may find your pairs strategy has a positive expected outcome but after applying costs, your edge is eroded.

@ Michael: thanks! Whether we should follow KISS always is subjective. There is no reason why a simple robust approach would always be better than, say, a sophisticated robust approach. In my experience, I have seen people curve-fit even using simple (in terms of understandability) technical indicators!

(personally I believe simple things don't work -- just my personal view)

On a different note, I agree that the back-testability of an EOD strategy makes it appealing/popular over an intra-day strategy. However, one question that comes to mind is that why should mis-pricing (deviations) between prices persist for long time-scales, i.e. over days, as compared to, say, a few minutes!? Perhaps mis-pricing is a misnomer!

Also, I agree with you and Dee that in intra-day there may not be enough gain after paying t-costs, but perhaps optimal order placement would help here, which makes the problem quite interesting to me.

@Justinas: True that unless I have the fastest IT, I may not want to trade something that converges within or less than a second, which is why I mentioned that I'd like to be able to filter out opportunities based on their estimated time-scale for convergence. So in this regard, speed is just one type of edge one might have; model edge and controlling the output of the application of that model (strategy, or what you call smart application of stats) are also very important in my view, and offer scope for investigation and improvisation.

@Ahsan: thanks! Let's catch up sometime time. Will add your id.

Observation : when you trade pairs [equities] you are trading two instruments that are in the same sector - industry - region = example : ANZ and NAB [ASX - banks]

= what you get are swings in the price differential in the order +/- $5.00 over 20 - 60 days

= so for a pair of banks whose prices move the same way just at different speeds this is a nice little earner - trade them with leverage [CFDs] and all you have to do is play these two instruments or include the other two majors such as WBC and CBA or WOW and WES [supermarkets] - nice little ecosystem

= means your analysis on the fundamentals is focused on 2 - 6 instruments and simply a couple of parameters that create the edge in 1 over the other instrument [there is always a poor performer that goes through patches to be a better performer]

= pretty clear price differential trends

= with a holding period of 10 - 20 days and getting $1.00 up to $5.00 + you dont have to split hairs on entry or exit or speed or cost

= within the large wave counts you get mean reverting waves so you can ride the smaller waves in the bigger with shorter holding periods

= so the question 4 me is - why would you trade intra day [you would if that is your style]

= and if you are technical orientated you can use MA or overbought-oversold indicators : the standard technicals that normally dont work on outright directional trades [I dont use technicals so I cannot vouch for these systems]

@Asim: you understood me correctly and you are surely right. If you consider a day with some simulated results inside, those would be a bit biased indeed. To achieve a good level of preciseness, it's better to obtain frequent data (such 5 min steps), then inside each step you can simulate a price with enough level of confidence and your result is going to be unbiased as more frequent data you use.

It does not really matter if you have EOD or intraday Pair strategies as long as you have a portfolio of systems. Then you will a more smoothed curve. Thats important about risk and leverage.

Pairs trading is another form of reversion to mean trading

- your trading program will need to determine the current 'state' the instruments are in = persistent LONG or SHORT trending, mean reverting LONG or SHORT and random

= as you know 'pair' trading is simply the reversion of the price differential - and it is this 'state' you need to understand

- it is simpler for single instruments to understand their 'state', but for pairs you can have two different states running and so this will impact the price differential direction

stay away-clear of random states : 4 mean reversion [persistent] risk management is around turnover and 4 persistent trend is 'go 4 it' [although you don't want to be going long in a short trend

So back testing requires knowledge of there states as well - to know what works - but if you cannot tell the states the instrument is in now, then you are playing with a loaded handgun

It is always the same and always will be - reliance on back testing alone to move forward is tough

This would apply for intraday as well as longer holding periods

and as to the 'random' state - it is very difficult to know when a price is 'random' - or who knows truly when they see a random state in price action

- again it would for the majority be the baseline assumption - but random prices can appear like trends

= so another issue to deal with in back testing

Richard G - are you using any statistical or quant methods to identify mean reversion? Sounds like you are just looking at the ratio of the pair's prices and guessing?

huh = no - 4 years of development and now about to release web application trading tools targeting CFD traders [newbies, semi experienced, intraday traders and active managers] so in Beta testing

managedleverage.com]

- the core tools detect the current' state' of each instrument : 'persistent Long or Short' [early, intermediate and late cycle], 'random' and' persistent mean reverting' Long or Short + the number of days [time] left in the trend

= but the main tool detects when the 'state' of each instrument will change before it occurs - say persistent Long late cycle will become random in 2 days and mean reverting persistent long in 5 days time - or the instrument is a short in 3 days time with a duration of 9 days in the short cycle

- and you can see this on a graph - the forward high/ low/ close and random + worst price is projected 20 days forward

- intraday [15, 30, 60 minute to EOD time cycles]

So 10 years research - 4 years software build : is it a guess = [maybe educated] - but it covers all instruments all assets including volatility and bitcoin and even asset allocation ETFs - through all instrument 'states'

- it was designed for new traders to CFDs or margin FX who are the 99% who lose their money and the broker platform is STP_DMA = close down market makers

but this is not to advertise - pairs trading is a new tool I am developing and saw this blog-thread = so a stock broker here in Sydney uses pairs trading in major stocks = financials, resources, consumer durables and so had to apply the algorithms to pairs - so what it does is tell you in advance -say 2-3 days when to buy a pairs = say buy WBC v NAB and the number of time units in the trend - say 9 days but you get to see the price path for the next 20 days and the series of states the instruments [pair price differential] will be in - so I have developed the tool and in specification with my developers

So walk forward - keep on walking - but to your questions

- I suggest the random price is random with drift

- and it is quantitative : 3 algorithms that work on i) single instruments and ii) indices + looking at iii) chaotic behaviour

Dee

How interesting - you work for a MM in CFDs

Yup. I used to trade for hedge funds and a prop house. Came back to the retail space to offer hedge fund strategies. It's working well for me this far. I hope you have the same success.

Thanks - good luck with that too - suggest educating the retail trader is the challenge - but pairs trading is a neat solution

- I used to be a professional institutional prop trader 20 years then started hedge fund then got into Risk with CFD brokers - margin equity and as well as profiling traders and designing optimal hedge programs

- the methodology of the tools came out of the CFD experience as saw the 99% lose and lose badly through GFC - so this the tools to know the road ahead but the trouble is newbies dont know what they need to know and sophisticated tools are bewildering [not to sound but sounding patronising]

==> the odds are seriously stacked against the newbie and the 'education' to date doesnt help - that is technical analysis and all the other toys, so if you want leverage you have to play the trade very differently and using technical analysis is not enough

But this thread is about pairs trading - so the interesting question for me now is = if the price differential is in a persistent trend [long or short] on the 24 hour time scale does it mean it should be on the same trend intra day say down to the 15 minute [equities] and maybe 60 minute for 24 hour trade cycles [with FX]

= so is 'time' the necessary component for 'persistence' - does the price need enough time to play out to create a trend as opposed to mean reverting or random and what is that time gap

- I have used 60 minute signals to determine the state of the instruments in the trade and use 15 minute to set the entry / exit but I find it difficult on any time line shorter than 60 minutes to get price range that makes sense [maybe a stretch at 30 minutes] to trade other than the random or mean reverting [without drift] but that is just me and you wouldnt trade these states even if you knew them

so if the piece of the puzzle to know the road is the price path over X units of time then what is the X units given the chosen methodologies

I'm not really sure what you mean. If you have determined a universe of stock pairs through statistical analysis (correlation, co integration and so forth) then you can rely on the observed behaviour. For pairs that exhibit mean reversion, there are only 2 states: diverging from the mean or reverting to the mean. Your identification of other states sounds elaborate and possibly meaningless. Good luck with your software.

Most interesting

= observed behaviour is historical not forward looking - it is in the judging the action of forward prices that make you or lose you the $profit and that is why you need to know the state and the duration of that state [even if the only state is reverting to or from the mean]

''For pairs that exhibit mean reversion, there are only 2 states: diverging from the mean or reverting to the mean'' - yup sounds fine

'"Your identification of other states sounds elaborate and possibly meaningless"

==> so if I understand your unique perspective, the only state an instrument can be in is "mean reverting" - not in a "random state", nor "trending state" as these are purely meaningless, at least in how you interpret the market states [fair enough]

- is it not possible that the price action of a mean reverting state is different to a random state price action as well as a trending state !

- so a crash is simply a big reversion to mean [if you look at it from several decades in price growth]

- so when a market price trends it is simply mean reverting, but unbounded, as a mean reverting state price action can certainly move beyond any of the previous boundary conditions

= so a trending market is simply a form of mean reversion, either trending away from the mean or reverting to the mean - ya that makes sense

As to random price movements - yes I can see how this too is simply a form of mean reversion as well - just very short time horizons [but then again who knows what a random price movement is]

- but if I am to understand this perspective, then is it as logical to simply state a market is "trending" either upwards in price [away from the mean] then down in price [towards the mean and then beyond] - there will be a mean value but is this meaningless as it is not static but moves about bit like the spot price does

- as well, a "random" price movement can be a series of trending markets [all you need is a price to move +/- 2 pips and you have a trend]

So how confusing - so we can now have a mean reverting state, a trending state and a random state and they all act the same, look the same, walk the same - so you say mean reverting and I say trending : you say mean reverting and I say random

but lets take your statement about anything other than "mean reverting" states to an even more practical conclusion : all states are meaningless [not just random and trending] as all you need to know is the target price whether that is 10 pips or 50 pips away [all technical trading systems tell you that] = who cares if it is a "random" state now or a "mean reverting" state or a "trending" state when all you want is to reach you target price - but in the end it is time, all you need is time and you will get there

- and each state has a different time component - given states are meaningless it doesnt matter that a trending market will get to that price faster than mean reverting and random would not want to wait for 50 pips in a random state [but then again it doesnt matter]

- no wonder most retail traders lose their equity through time and it is the Market Makers who win as how would you know what state you are in [oh yes, it is a mean reverting state - historically speaking, but it may be a mean reverting state diverging from the mean or reverting to the mean in the next price move, and what about the next and the next and the next price movement [of yes all mean reverting states]

Dee - how do you reconcile this [I have worked in 3 CFD MM and I know their business model as I profiled traders and designed optimal hedges] - you advise retail traders in CFDs in a Market Maker who make their money from client losses ?? - so why are you there if your clients are making money

So for definitions : just to focus the discussion

trending price is defined : probability the next price's direction will be in the same direction as the last price [the higher the probability the stronger the trend]

random - 50 : 50 probability of the directional price move

mean reverting : probability that the next price direction is in the opposite from the last price move [mean reverting can be long or short trending]

You can get degree of probability - the higher the probability the stronger the state [persistent, mean reverting, random]

So if you knew these states : "persistent long / short", "random" and "mean reverting long / short" and the time of the trend, you have the projected price path - EOD or intraday price signals

So all you are doing is playing the probabilities - if you don't know the probabilities everything may appear as random

A guy named Didier Sornette has been doing this since 1996