Commodity Trading

Case Study

Market Situation

Quantitative Finance is the trendsetter in analytics, using advanced algorithms and speed of data to profit in the global marketplace. Modern Analytics suggested a key differentiation is the ability to access far more “big data”, new, richer sources of information and vastly increased throughput in the number of algorithms deployed.

Background

The firm trades various commodities on the Chicago Mercantile Exchange using machine learning technology and advanced analytics to drive decisions. They have automated the entire trading process by having Modern Analytics develop an application using SAS to create trading rules and execute the resulting trades on a millisecond basis.

Objective
  • Create a data warehouse with level I&II commodities data (~100GB/commodity) and use it to predict short-term (intra-day) market moves.
  • The ultimate goal was to find proper short-term entry and exit points for long and short term positions – without any human intervention. To efficiently and effectively roll out and maintain the system across a wide range of commodities, model design needed to reflect only a few highly selective parameters.
Strategy
  • The strategy was to apply proven advanced statistical methods to evaluate second-by-second opportunities in the commodities (futures) market.
  • Also sought out, was a closed loop system where at the end of the trading day, the historic data could be retrieved and used to validate the system’s intra-day trade executions. Some of the reporting produced compared the “lab” to the “executed” intraday trades allowing for an assessment of the slippage and percent overlap between actual and expected trades.
Solution

The self-learning system created utilizes socket-level data feeding off exchanges (such as the CME) to score a hybrid structure of 15 regression and 15 neural network models. By performing a fuzzy logic signal evaluation the models were capable of placing automatic trades (short and long), all while monitoring the days position from a risk and profit perspective.

The underlying data sets per commodity consist of 5.5 million tradable seconds per year and are roughly 5,000 analytical variables wide. The system utilizes parallel processing to create statistical models from scratch and scores these models to provide forecasts with variable targets and inputs. Currently the system is using 20 minutes of history (inputs) to predict two minutes into the future (targets) for a variety of commodities (futures) such as ES, NQ, and ZN to name a few.

Results
  • This self-learning system needs almost no human supervision, is extremely efficient in finding
    optimized entry and exit points– while doing so every second of every day for as many commodities (futures) as needed.