I appreciate your comments Nitro. Let me try to frame what is probably ineffctive thinking. I am aware of the highly directional nature of long/short high - corr assets (say, -.8 to -1) seeking trend. Conceptually , such a spread is really just a couple outrights. I do remember Bwolinsky advisng readers to select very high + OR - correlations to spread (MR or relative out performance where daily net of vol is less than each individual leg in the highly + case AND pure trend in in the highly -corr case). Dont diddle in the middle in the land of uncorr. I tend to also treat the + corr spread such as AAPL/QQQ as two outrights AS WELL despite their tight +corr. because i look back 5 years and see the breakdowns and assume it will happen tomorrow and size as two -corr outrights but usually do GET the unexciting results of two +corr. Maybe i must decide on a 95 to 98% simple confidence analysis and accept that risk with better(larger) sizing. I spent some time today looking learning of lead/lag as i have simply not looked at this in the past. Keeping it simple i intend to collect data on the spreads i watch and measure led/lag. Intuition, familiarity may allow one to say it is probable that when x goes up then y will go up as well but i have not looked at delay/timing/phase (in days for me) . Not sure how to go about testing lead/lag. Any thoughts would help. Thanks. I have looked at tabular/matrix corr data that i have formattd as well as histograms of 30,60,90 Day corr. in charting package. Lead /lag though....no not yet.
Right, those are both good measures and vector autocorrection/autoregression etc are great ways to approach modeling markets. The problem is that most, for lack of a better word, unsophisticated mathematically, traders can't even get started with these sort of models. However, if you can approach your markets in this form, all sorts of very powerful tools all of a sudden apply. I love the function Code: casuality() If only it were so easy!
Great! How are your legs doing after Draghi and China IRs announcements? If you are losing or winning on all legs by relatively large ratios to your positions size, it is usually a sign that you are too directional. Or, that you have not modeled correlation or cointegration to take lead-lag into consideration. Notice that the context matters. So, if your model is able to tell that the market is not panicked (in either direction - here, the equity markets is quite close to melting up parabolic), and that the co-movement measures show that the model is still correctly updating its views, then adding to a losing position may be appropriate. That is the price we pay for mean reversion, and why it is so much harder to trade it than momentum. And, we haven't even started talking about behavioral economics as it applies to trading, which is probably 25% of any good system. I use it mostly for timing. Classic charts are decent at capturing that. But how do you recognize/capture that in a model? These models sound like have a mean-reversion feel to them, but they need not be. If all of this sounds complicated, it is because, it is!
Haha. I want that function nitro. So if a trader were to educate himself on lead lag, where should he start? Is there a text you recommend?
Well, I think you are already on a (notice I didn't say THE because it is a moving target) good course. If you are used to R, start with that and follow your nose. BTW, it is a huge undertaking. Getting from a trade-able model to being able to trade it is not simple. But at least hopefully the people that read this get a feel for the way that HFTs and other institutions trade. Leave the retail world behind, and demand from brokers the same advantages that institutions have. One thing that has gotten fairer is the tech, software, hardware, and access to exchanges. So much of this stuff was even five years ago totally out of reach for the retail trader. Now, there is a pin prick of light coming from the end of the tunnel. In some sense, all of this type of trading comes down to being able to tweak a set of knobs, so that the model responds in a certain way. The greatest lesson I have ever learned I learned from an absolutely great options market making firm. It was not the model that mattered so much, it was the way the model was so amenable to tuning the knobs and that those knobs were very precise in responding to market conditions and what the trader expected from that tuning. I know it sounds ridiculous, but I have never seen anything like it. Let me give you an example. Take the crack spread. Say you wanted to model it. Your models success will come down to, what is the right ratio!! (I am oversimplifying in that in this case, you already found related instruments to remove direction and are cointegrated) Is it 1:2:3, is it 5:3:2 etc? Sounds absurd, but what is lying deep underneath that is a model that allows you to very precisely turn knobs such that those are the ratios that it returns. Then you have invertibility, so, you can go from ratios to model behavior, or model behavior to ratios. You still have to drive the car around the track at 220 MPH, but now you have a racing machine that responds exactly the way you expect on straightaways, corners, breaking, accelerating, etc. I hope that makes sense.
Sure it is. But let me step back and mention where I am at in my own current path. Much has been written about Finding Fair Value (FV) of a set of instrument(s) Controlling Risk as ratios of legs put on. "Portfolio Management". Optimizing position sizing (sizing the actual ratios themselves) so that the returns are optimal with lowest acceptable risk. So if the model returns a:b:c, what are the actual units, X,Y,Z such that the actual position is aX:bY:cZ. Kelly says you also have to worry about X,Y,Z, not just a,b,c. You see what I am getting at? Somehow, those three things are highly interelated, and yet, we treat 1 as say VECM, 2 as ratios, and 3 as Kelly Criterion (I am just using the generic term.) But what is the one unifying view of all of this? I can easily see that 1 and 2 should be part of the same model. I can see that 3 is external because it depends on your own wealth and risk tolerance. But is it really? I guess what I am proposing is that you approach it not from putting on the legs side, but from a unifying at least 1&2. Ideally what you do is, enter inputs, and out pops a model that is tailored to optimize wealth creation for a given set of instruments. It is very similar to compiling source code to machine code! But back to earth, look around at the R sites. There are lots of examples there on how to approach this sort of thing. Make sure that you integrate real costs of trading like commissions and if it is not a mm'ing system, the bid ask spread when doing your testing. There are very few if any turn key things that will work. If you don't know how to program, and how to read other people's code, it is going to be a hard uphill battle for you.