Questions:

• CTA strategies are popular, but often not suitable for a smaller set of markets, e.g. commodity sector alone.
• Can we create a ML strategy capturing the main performance attributes of traditional CTA strategies?
• Using the AiLA approach for trading strategy modelling.

Approach:

• Replicate Momentum, Carry and Value CTA signals for ML training, based on public literature [Baz15].
• Construct a long/short signal prediction model as described in the CTA ML modelling section.
• Compare performance between traditional CTA signals vs ML signals, for 29 commodity markets typically included in indices like BCOM and GSCI.

[Baz15] J.Baz et al., Dissecting Investment Strategies in the Cross Section and Time Series (December 4, 2015). Available at SSRN:https://ssrn.com/abstract=2695101

• For clear comparison, CTA and CML (CTA-ML) signals are produced in equivalent format and applied to the same market returns.
• A performance difference therefore directly relates to the signals, and portfolios are compared with same daily risk.
• A daily P&L correlation of 58% is obtained between the CML and CTA portfolios, indicating a substantial similarity with the performance from traditional CTA strategies.
• At the same time, the CML portfolio indicate a significantly higher performance than the CTA strategies over the given history and on a risk adjusted basis.
• It should be noted that in average CML hold positions less often than the CTA strategies and all results are ignoring transaction costs, expected to affect CML results the most.

(* The returns used are based on the front month contract for each market, typically rolled a couple of weeks before the contract expiry)

• The role of timing is studied by masking out days when CML have non-zero positions for the individual markets.
• Applying the position mask to CTA signals, indicates an enhancement using CTA signals on days selected by CML.
• CML performance is broken down into contributions from days with different types of CTA signals, i.e. same/different to CML signal, or CTA signal being classified as zero.
• Most days with a CML position belong to the same signal category, also indicated by the CML vs CTA correlation.
• CML performance on days in same or zero CTA signal category indicate a sizable positive CML contribution.
• CML performance on days in the different CTA signal category indicate that CML is more often right than CTA, however, also a weaker performance when signals disagree.

(* For the comparison of CTA and CML signals, the net CTA signal on a given day is classified by belonging to either a long, short or zero signal group)