DeepState
Multi-Strategy, Quant and Fundamental
- Joined
- 30 March 2014
- Posts
- 1,615
- Reactions
- 81
Howard, it's not the broad process mechanics that I am questioning. Though, thanks for writing it out again, and in more depth, to shed more light.Hi Deepstate --
To my thought, the procedure is pretty clear and pretty much all trading system developers have the requisite domain knowledge.
... Development:
1. Use the data prospector to curate a list of potential candidates.
2. For each candidate, attempt to develop a long / flat trading system using the scientific method. The model is rule-based, uses state signals, marks-to-market daily, is looking for trades that are accurate, frequent, and hold a day or two. If a traditional trading system development platform is being used, an oscillator with a fast cycle works. Pick any one, since at that rate, all are equivalent. Or some other indicator of your choice that will give about 50 signals per year.
3. Also attempt to develop short / flat systems, but there will be lower success.
... Trading:
1. Bring data up to date and evaluate all the systems daily, computing safe-f and CAR25. If the one with the highest CAR25 has a score higher than a risk free account, take a position using safe-f portion of the account.
2. As the jingle says -- rinse and repeat daily.
Still one system at a time, one day at a time.
Volatility was one of the characteristics examined by the prospector. When he or she passed it originally, volatility was acceptable. If it increases at a later time, dynamic position sizing will identify that and lower safe-f and the CAR25 score, moving it down the list of those available to be traded.
If AmiBroker is being used, I have published several models that work.
If Python / machine learning is being used, I have published one decision tree model showing that it replicates the traditional platform. Also examples of about ten alternative models that can be applied with little change to anything. And suggestions for additional predictive variables.
Modeling experience and skill become more important here. The magic sauce, whose recipe developers of machine learning trading systems will hold close, is which additional predictor variables are useful, what steps are taken to create stationary data, and what data preparation is used.
Best regards, Howard
Ultimately, this is another process framework with a range of parameters upon which the success and failure of an effort will rest. In broad terms, I agree with the concepts used in risk deployment. There is merit in checking for performance in a recent period in the belief that regime rigidity is a feature and might help with prediction in that environment. CAR25 is a selection criteria and is a reasonable one.
None of these will save you from throwing garbage at it. And, if the skill of the operator is what differentiates the winners from losers, it was so before this approach was conceived and remains so.
The main question I have relates to portfolio construction. This is a thread where a strong statement about 2 stock portfolios was raised. In my view and in my experience of seeing and participating deeply in professional money management of the real and hedge variety, this is unsupportable for situations outside of play money. And that's fine if so.
Yet, you came along and seriously expressed support for a 1 stock portfolio. And it is this I wanted to explore. Your proposition is 1 stock at any given time consuming the full risk allocation. You will appreciate that the position is an extreme one, but you are a man of serious intellect where statements seem to have rationale, even if we might reasonably disagree on those merits.
So, I say to you that, even under the framework of CAR25 and safe-f, you would be better off holding more than 1 stock at a time. Yes, your current process only selects the best CAR25 over some horizon using all sorts of permutations along the way as may be codified. End of the day, Howard Bandy: 1 stock...full risk deployment on safe-f.
I do not think I am misunderstanding you when I say:
+ Surely, if you had a million well chosen signals feeding in to a soup of algos over a range of instruments, and their merits, however obtained, were ranked first to last on any given day...that you would feel that only one was valid. Yet, that is what the system does.
+ If signals other than the one ranked one have any merit given the skills of the operator, others with lower rankings also have solid expectations. Let's say that these signals are issuing a BUY on some other instrument so there is no overlap. These are currently not used in the Bandy System.
+ By not using signals which also have solid expectations and allowing for portfolio effects, you will produce worse risk adjusted results than otherwise achievable. If the operator is skilled at selecting candidate signals for ranking, that would be a huge information loss.
+ By putting all weight on the 'best' only, the system actually produces a risk-adjusted outcome which is far from best in a portfolio sense.
Reasons why you would not do this include:
+ A trivial case of where the inputs are known to be garbage and no algo to screen them can convert garbage to alpha without finding a latent factor via the screening alone.
+ You literally believe only one of the candidate signals in the primordial soup truly had positive expectation. This would justify full risk deployment on this one instrument on the one signal on this particular day. Whist things may be non-linear in life and require special techniques to identify the underlying order, this kind of binary, first is right and the rest is junk for today, outcome is wildly unreasonable in my view or a sign of a highly unstable set of candidates where identification of merit would already be highly tenuous.
Comment sought:
1. Do you think that only one signal, that chosen as 'best' by the process for a given day, has merit in terms of positive expectations? That the rest have absolutely no merit at all on the ranking criteria beyond that placed first?
2. If the screening criteria does rank in a way which helps us order the merit of signals and others are also somewhat valid, why should this information not be used and result in these signals combining in to a portfolio setting, given an aggregate safe-r equivalent (adjusted as required for whatever you like to get this estimate) and then deployed in to the market. More than 1 stock at a time.
Regards
DS
Last edited: