Having previously discussed about types of Popular Systematic Strategies, we shall try and address the question of how does one go about actually designing one. Specifically, we shall look at the various elements of a strategy and the considerations within each. Bear in mind that the order of steps listed here is generalized and some of the steps are interchangeable as per requirement.
Step 1: Hypothesis generation
As with any other area of research, an idea is where any strategy starts, whether systematic or discretionary. The only difference in a systematic strategy being that one would actually want to verify the validity of the hypothesis using historical data or simulated data using historical variables. A hypothesis is simply an idea that needs to be verified whether it holds true with actual data. In the context of a trading strategy this could be the expected behaviour of a security or portfolio when a particular set of conditions are met. For e.g. stocks with lower P/E outperform stocks with higher P/E across longer periods. The source of ideas could be market observation, books, data analysis, forums or just a simple thought exercise. Occasionally, steps 2 and 3 might be a part of the hypothesis itself.
Step 2: Universe identification
Once a hypothesis is generated, we need to identify a suitable target universe to test it on. This step is intrinsically linked with the hypothesis in that the choice of the universe largely depends on the kind of idea we have. For instance, if the hypothesis uses earnings data or analyst estimates as a condition, the target universe has to be equities. The universe could be a broad asset class like equities, debt, commodities etc. or could have further classifications like large cap developed market growth stocks or emerging market investment grade corporate debt and so on. Generally speaking, the smaller and less liquid the universe is, the more the likelihood of successful verification. A good way to validate this is to start with a smaller universe and then cross validate with a larger universe with similar characteristics.
Step 3: Data
Probably the most critical component of all, data includes input data choices (what are we looking at to make a decision), output data choices (what are we trying to see an impact on) and all intermediate processing and cleaning steps. While some hypotheses can be very specific on the kind of input/output data, most will be fairly generic and can be implemented using multiple different metrics. For instance, if the hypothesis is that cyclical stocks do well during economic expansions, the individual definition of cyclical stocks and economic expansion need further definitions and ‘do well’ itself needs a time horizon and further description. Different implementations of the same hypothesis can lead to different observations and requires a deeper investigation. Cleaning and processing of data are equally important steps, as unreliable or poor data can lead to erroneous conclusions which can be disastrous.
Step 4: Capital allocation
Once a target universe and input metrics are finalized, what remains is to assign relative weights to each individual security or to assign position sizes for each individual signal (buy/sell). If the strategy is a portfolio based one, we would need to decide relative weights of each security in the portfolio and if it happens to be independently run on multiple securities, then the relative size for each decision needs to be quantified. For e.g. a factor based long portfolio could be constructed using market cap weights, equal weights or volatility weights. Similarly for a directional strategy we could have target positions sized according to the volatility of each security or at the time of trade. The choice of capital allocation often serves as an enhancement of the performance of the core idea.
Step 5: Historical analysis
Commonly known as a ‘backtest’, this step involves testing the hypothesis using the data identified on the target universe. A standard analysis simply includes looking back at historically observed actual data while a more sophisticated version could look at simulated data using the statistical properties of actual data. This step involves evaluating and understanding various aspects of the hypothesis and often times researchers adapt the initial data based on the analysis; however, one needs to be careful not to overdo this or to do it on a control set so as to avoid data mining/optimization. Once completed, one can either reject or accept the hypothesis or further iterations of the above steps might be required.
Step 6: Risk control
Risk control or mitigation is often built into the strategy to ensure that adverse events are manageable and within the expected limits of the strategy. Stop losses, concentration limits, position limits and liquidity-based allocation are common measures that are used to mitigate risk, occasionally at the cost of some performance. However, this is an essential component of any real-world scalable strategy as an unanticipated event could wipe out a long history of good performance. Some strategies might use derivatives to hedge some unwanted risks or to limit worst case scenarios. One must always be prudent while designing risk controls as a lot of risk factors get overlooked simply because they are not very frequent.
Step 7: Practical considerations
There are a few other considerations that could impact real world performance of strategies such as choice of execution logic, transaction costs and managing implicit costs such as slippages. The choice of execution logic depends on the nature and frequency of the strategy as well as the liquidity of the underlying security; for e.g. a large position on a stock with low liquidity will need to be executed in a very different way than a small position on a liquid index derivative. Similarly the choice of order types (limit orders or market orders) will determine the percentage of orders executed and the market impact of our orders. Transaction costs can be a huge dampener for strategies that trade a lot and can result in a significant drag, if not accounted for. One must always ensure that these are factored into any analysis before deploying real money. Lastly slippages, the difference between expected and actual execution prices can be unknown and modelling these into a strategy can be a separate challenge. However, prior estimates based on other strategies or based on a short actual market implementation can be helpful here.
The above mentioned steps are in no way exhaustive or exclusive and the implementation and ordering of each varies between different practitioners. While a lot of the above can be automated and is scientific, the art of designing a successful strategy lies in combining all the various pieces into an effective whole.