Journal entry (Wall 'o text) Previous entry Education I am up to lecture 9 in the Quantopian lecture series. I also learned that Quantopian will be adding futures to the lineup. This is probably good news. Thoughts on my system I've been reading through some of the posts here and I've seen a lot of the following: Traders who can't elucidate their strategy and when they do, it sounds like mumbo jumbo. Traders who have a plan and stick to it. Discretion is a last resort. Traders who want to call tops and bottoms. Traders who look for mispricings. There are probably more types, but these stand out to me. The ones I probably feel more of a kinship to are types 2 and 4. How does this inform my system? Well, for one I won't develop a strategy that lets me make any decision. Period. Finding mispricings One of the things that sticks out to me way back when I was writing code to value MBS was when I said to one of the quants "hey, what happens if this tranche is full of total bullshit?" "That is unlikely to happen." "Well, you have a PhD, you must know better than me." That always stuck out to me. I think I have an intuition for smelling out bullshit and that could probably be where mispricings can be found but as a retail trader, it is laughable that I would be able to access the kind of mispricings that Zzzz1(?) had located where he could arb between dealers who had no idea how to price the contracts that they were selling. As a retail trader, the most complex contracts available to me are futures and options. Complexity is where profit can be made because that is where people screw up their valuations. But as available as they are to me, these are well-known contracts so it is impossible that they will be mispriced. See for example, this post here about EBAY %IV on April 21 options and how quickly the market adjusted when the earnings had been pulled up. Mispricings can be located in different ways. If measures A & B are beginning to show correlation but the market has priced them to be uncorrelated, that could be an arbitrage opportunity. It can't be obvious things like price. For example, perhaps the number of tampons sold never used to affect the price of oil, but now that women are working more, they need more tampons to keep stocked at work and so to get to work, they will need more oil. Obviously, this is a facetious example but this is the kind of correlation I am talking about looking for. The point is: the easy mispricings will be unavailable. There are people who will have better information and access than me. Am I missing something here? Using Beta to make profit This leads me to the conclusion that the first strategy I should aim for is to use beta (i.e., market movements) to make a profit. One of the things I noticed the first time around is that there are regularly times when there is great volatility in the market: opening, closing and around news. I've already concluded that I do not want to play news because at the moment, I do not have any idea on how I could systematically do so. I have also concluded that "baby's first algo" should be simple. I saw this post earlier today about playing the open which confirmed exactly what I had been thinking. Over the next week, I'll create a notebook in Quantopian and mess around with it to see what kind of backtest I can create along these lines. Until next time...
A few comments -- I highly recommend iPython notebook as a research tool, the whole embedded flow is almost ideal for strategy development -- don't think about the intricacies of the execution at the stage of strategy development, that's next step
I assume you don't mean intramarket pricing within a single product. The futures markets are painfully efficient as simply getting the bid/ask spread reliably will make you a very wealthy person. People look to regulatory and exchange pricing mechanics to arb these out within hours(maybe minutes?). A cool project: Perhaps, you can write a basic machine learning algo to curve-fit the prices of various contract months to a minimum R^2. (Note: Self-serving interest.) The other side of the spectrum is essentially an allocation decision based on some relative value model between different asset classes or even different types of premia. These strategies are more accessible with your platform. I would use target premia to classify my strategies. This will capture the attention of an audience that ranges from academic to professional. Also, go with the flow with what Quantopian offers you. They are techies with an asset management heart. If you use technical patterns or price action, your work will most likely be buried in obscurity. Or, you become incredibly wealthy.
That is a great suggestion regarding IPython for research. Is the reason you're pointing out to ignore execution at this point because Quantopian probably isn't the platform I want to use for that?
Thanks for reading, so many things to consider in your post. That seems achievable. Can you tell me more? Do you mean that people analyze regulatory behaviour that the market would not be aware of? Why could you not do this using some sort of non-linear optimization with the objective function to minimize being R^2? What is it that you think you need machine learning for here (and what algorithms?) Is it just because the relationships are so non-linear that you have no hope of using any statistical algorithm with success? I'm perfectly willing to try, but I need more info I don't understand what you mean by target premia. Can you give an example? Do you mean "I want to target $2.50 per trade" or something to that effect?
If you are proficient in programming and ready to sift thru exchange documentation. Maybe. Read @cjbuckley4 threads. The person was working to build a formula one car engine out of his garage and got picked up by Renault. Nuff' said. You will be competing with every jaded IT programmer who has been coding C++ for the last 10 years and trying to break into a front desk job. As for regulatory, you may have to put a few glasses of scotch thru Bob Morse for a recollection of how regulation thru the years affected pricing/margining mechanics of stock options. I am no expert. You would simply be using a hypothesis/cost function that is logarithmic as opposed to linear. The objective function remains the same. Perhaps, it is not an efficient use of ML. But, I have programmed such an abomination. Additionally, I have no familiarity with any non-linear optimization methods. Hopefully, you are teasing. This is a major knowledge deficiency which may loose us the attraction of some ET's finest to this thread. Some of my mentors, idols and advisors. Volatility risk premium, carry, credit risk,large cap vs small cap to name a few. Most, if not all, expressed in many different ways, in many different products and many different trade mechanics. You could start off with an asset class/sector rotation strategy. This is the equivalent of the MA crossover system for technical analysis based algos. This is Everest as execution is K2. Which summit shall you make first?
I think you misunderstood my comments on broker arb. All those contracts per the most basic index options, some var swaps. The price discrepancy purely originated because of poor inputs, volatility in my case. We talked about pricing models and I elaborated and stated that a lot (not all) of pricing models are pretty sound and solid and that it is the inputs and assumptions of inputs get us in trouble every now and then. Just wanted to clarify this confusion
Can you elaborate a bit on the python research effort? I have to admit I am a bit of a burnt child when it comes to scripting languages, I seriously considered R and Python many years ago to promote as primary research tools. I did not like them for various reasons and decided to create my own profiling and testing platform. I never looked back and instead constantly expanded on features based on need. It even allows me via Roslyn script engine to write and execute a script during runtime, cleanse, prune, visualize the data, generate any desirable stats,... I see the point in using Python to get up and running quickly. But for someone whose life (and the one of his wife) depends on generating ideas, profiling them, testing them, and translating those ideas into strategy code and ultimately profitable trades, I just saw the need for a more rigorous and comprehensive platform than stitching together tons of different R and Python libraries and hoping the source code under the hood does not contain errors and not knowing whether support for updates will continue or not. It just never appealed to me. Ironically, I still take a look at what Python and the core analytical libraries have to offer every now and then...
Thanks for the clarity. Did you ask them what models they were using to price? I'm curious how you determined that they were using bad volatility inputs.