Systematic vs. Discretionary? The quantamental future of trading
Updated: Apr 18, 2021
Historically, Particle.One has been working with buy-side customers, offering data organized with the knowledge graph methodology and tools to accelerate hypothesis testing.
Our cloud-based AutoML platform makes sound quantitative decision making easy and efficient.
This year, we started working with physical traders and large commodity producers, helping teams that manage risk understand price movements better. Among the early technology adopters was Frey Commodities – a physical trading company that believes in data-driven decision-making.
The flows of information and commodities are in many ways disconnected, and the physical trading is characterized by many inefficiencies and bottlenecks that are all dynamic. As such the physical aspect adds multiple layers of complexity in terms of trade execution, but also offers an opportunity to collect data and express views on the market. The aim should be to make better informed decisions not only at the time of trading but also at the operational level when planning and executing the physical flows Jeff Løcke Laursen
Particle.One AutoML is an interactive quant research tool that empowers users to build trading models without requiring in-house quant expertise. It allows running data analysis at scale and testing the predictive power of features over time, and selects the most important predictors from a set of candidates.
A huge question for the discretionary market and risk management is, How accurate can systematic predictions actually be?
To generate healthy returns at an annual scale, a systematic strategy needs to be directionally right just slightly more than 50% of the time. The benchmark of human traders making price forecasts is considered to be significantly higher. Discretionary trades are placed at much lower frequencies, pushing the need for the confidence level behind the trade to be higher.
The intersection of systematic and discretionary trading is the quantamental approach to building models.
A use case our team pursued at the request of a customer was to understand the dynamics of local basis prices for the soybean market in Kansas. The trading horizon of interest was 8 weeks.
In this paper, we only use publicly available data to build the price forecast.
The discretionary team working with Particle.One AutoML tools wanted to construct a model forecasting binary directional price movements for Kansas basis prices. The economic drivers behind the price movements were a combination of fundamental factors and proprietary data (which enhanced the performance of the model but is excluded from this white paper).
Basis price is defined as the difference between the cash price paid for supplier’s grain and the Chicago Board of Trade futures price.
The initial step for any quant work is to understand the nature of the data that will be used by the system. How complete is your data? How long has it been systematically collected? How does it compare to alternative sources or proxies? These are a few of many important questions to answer in considering a data source.
Publicly available data at Agmanager.Info contains weekly data for different providers. Cargill price data was selected based on internal criteria.
How complete is your data?
How long has it been systematically collected?
How does it compare to alternative sources or proxies?
Price series are not good candidates for out-of-the-box prediction methods. The data needs to be first transformed into a stationary series (e.g., looking at returns derived from prices is a good first step). In the case below, we calculated a z-scored return of a shifted price series.
Various economically-motivated factors were considered as potential model features:
Month of the year (to capture seasonality)
Closest futures contract expiries
Substitute prices/returns (e.g. corn)
Soybean crop condition and progress
Transportation costs (rail tariffs, transportation cost indices, etc.)
USA soybean export/production
The features with the largest correlation with the target were:
November and January expiries (indicates whether the closest expired futures contract is November or January)
The share of the crop in good and excellent condition today relative to the 5-year average value
AutoML suggested a model taking into account:
price history (autoregressive/moving average)
economically motivated predictors
It is updated once a week and forecasts prices 8 weeks ahead. The same model can be used to generate forecasts at shorter and longer time horizons (8 weeks was chosen beforehand by the partner).
The results of the model performance are highlighted below:
We use Sharpe Ratio as a measurement of model performance. In this particular case, the intent was to understand the directional price movements better. Our model was able to achieve an average 76% hit rate for when forecasting prices 8 weeks out.
AutoML tools enable physical traders explain price movements in a data-driven way. Using this technology one can build similar models without having to hire a large quant systematic trading team.
To try AutoML tools and schedule a technology demo, contact email@example.com.