Plunging into uncertainty
Building an Alpha Model with Senticon
Adapting Senticon for Risk Models
Trading with Senticon
“Optionality is Promethean, narrative is Epimethean.”
N. N. Taleb
We humans are creatures of narrative. This serves us well in the string of interactions that our life - itself a narrative - consists of. However, place us in an environment characterized by coalescence of millions of individual narratives, and our neural predispositions are quickly relocated next to our appendix at the corner of Vestigial and Curious.
When immersed in emergent phenomena, whether they be prices on a ticker or hashtags on twitter, our ability to predict is almost completely compromised. Simply put, we’re just not wired to think in aggregates. We abstract the aggregates and create a higher-order emergent system, one that is characterized by a very small number of newer, emergent narratives, and attempt to attack the problem of prediction with this newer, larger, and diluted tool.
How good is our “new” story, our abstracted model of an emergent system? Let us consider a framework to evaluate predictive models in our far-too-often inconveniently uncooperative reality. At the very least, this framework will hopefully help frame conversations around the swarmsignals content.
“What did I think would happen?”
Jim Lebenthal, Bernice Kanner
The objective of any model is to predict some aspect of the system it is modeling. It takes in some set of data, and outputs a prediction. As no model can be completely accurate, some level of error is expected.
Alpha models can be either theory driven or data driven. Theory driven models detail what should happen based on behavioral assumptions on how outputs of the model relate to inputs. Data driven models rely on techniques such as data mining and other statistical methods to determine behavioral relationships between inputs and outputs, and assume those relationships will hold going forward.
Here at swarmsignals, we will initially be focusing on data driven models. However, building theory driven models is not entirely out of scope. Consider, for example, a model that predicts next quarter’s performance based on the levels of stress detected in the voices of the conference call for this quarter. Or, perhaps, a model that monitors cellphone traffic at various locations and attempts to derive macroeconomic assumptions as well as retail traffic.
The understanding of what kind of model we are working with is important, when attempting to use tools such as correlation and multi-parameter optimization to backtests. Additionally, the type of model also has implications on what parameters are allowed to vary: a data driven model, for example, may allow significant variance of the model parameters themselves, as no narrative constrains our model; however, a theory driven model may not be modified beyond the bounds of the narrative imposed by the theory.
“If a man is both wise and lucky, he will not make the same mistake twice. But he will make any one of the ten thousand brothers or cousins of the original. The mistake family is so large that there is always one of them around when you want to see what you can do in the fool-play line.”
Risk models represent the nemesis generated by any position taken in response to an alpha model. They attempt to quantify how much taking a position can lose. Risk models can quantify types of risk, amount of risk, or both.
In the last two decades, a lot of work has been focused on characterizing risk via various models, ranging from risk incurred by a single position (VaR, Extreme Value Theory, etc.) to risk at a systemic level (NYU’s V-Lab’s various methods).
From a financial trading perspective, trades are ideally structure so that the risk/reward is extremely asymmetric, such that our exposure allows for the most gain with the least risk. Operationally, Jamie Mai of Cornwall Capital and Michael Burry of Scion Capital are excellent examples of how this can be done.
It is important to note that risk models represent more than just “stop” or “max pain” levels. If the trade is designed to be held overnight, risk models must incorporate gap risk, perhaps using tools such Average True Range (ATR).
“Guessing what the pitcher is going to throw is 80% of being a successful hitter. The other 20% is just execution.”
Once we have understood what we expect to happen (alpha) and the risk we incur by taking a position (risk), all that remains is to take the position. Execution models quantify the details in actually taking the position.
Execution models may include transaction costs (although frequently, transaction costs are given their own model). They usually include the trading venue for the instrument in question, positioning strategy, and order types.
Execution models also frequently include the cost of exiting the position, for both unfavorable or favorable exit conditions.
“Life is a series of natural and spontaneous [portfolio choices]. Don’t resist them; that only creates sorrow. Let reality be reality.”
Lao Tzu (paraphrased)
The metrics resulting from the alpha, risk, and execution models described above can be used to make a decision of whether to take a particular position or not. However, once this decision has been made, the question of asset allocation arises. What to allocate, and how often to change the allocation, are decisions handled at the portfolio level.
Broadly speaking, there are rule-based portfolio models (such as Ed Throp’s Kelly-criteria based models, or models that rely on each strategy’s equity curve), and optimizer models. Optimizer models are a field under heavy research currently, especially due to complexities involved with calculating the risk incurred over the entire portfolio.
Portfolio models are often subject to transaction costs as the portfolio is rebalanced to address the changing risk profile as it evolves with time. For example, a modern delta-neutral model will adjust positions to accommodate the changing delta of its positions at varying time intervals. For an accurate understanding of how a set of strategies perform with a given portfolio model, these costs must be accounted for.
“If we knew what it was we were doing, it would not be called research, would it?”
With the above main models, we can start collecting performance statistics of our portfolio. These metrics can then feed back to our various models as desired.
The following image sums up the process of a single iteration.
The above framework represents just one of many possible frameworks for trade and portfolio structure, but it is one that should offer a good starting point for discussion on use of the data offered by swarmsignals.
As the saying goes, the way to get started is to quit talking and begin doing. To that end, I hope to tackle building trading models with senticon data in my next post.
The ludic fallacy is a term coined by Nassim Nicholas Taleb in his 2007 book The Black Swan. “Ludic” is from the Latin ludus, meaning “play, game, sport, pastime.” It is summarized as “the misuse of games to model real-life situations.” Taleb explains the fallacy as “basing studies of chance on the narrow world of games and dice.”
It is a central argument in the book and a rebuttal of the predictive mathematical models used to predict the future – as well as an attack on the idea of applying naïve and simplified statistical models in complex domains. According to Taleb, statistics works only in some domains like casinos in which the odds are visible and defined. Taleb’s argument centers on the idea that predictive models are based on platonified forms, gravitating towards mathematical purity and failing to take some key ideas into account:
In financial mathematics and financial risk management, value at risk (VaR) is a widely used risk measure of the risk of loss on a specific portfolio of financial assets. For a given portfolio, probability and time horizon, VaR is defined as a threshold value such that the probability that the mark-to-market loss on the portfolio over the given time horizon exceeds this value (assuming normal markets and no trading in the portfolio) is the given probability level.[clarification needed]
For example, if a portfolio of stocks has a one-day 5% VaR of $1 million, there is a 0.05 probability that the portfolio will fall in value by more than $1 million over a one day period if there is no trading. Informally, a loss of $1 million or more on this portfolio is expected on 1 day out of 20 days (because of 5% probability). A loss which exceeds the VaR threshold is termed a “VaR break.”
VaR has four main uses in finance: risk management, financial control, financial reporting and computing regulatory capital. VaR is sometimes used in non-financial applications as well.