• A common concern regarding strategies trading outright futures is often their similarity to CTA strategies, such as momentum, carry etc.
• Despite the diversification benefit with respect to other investment classes, CTA products tend to have high correlation between each other, e.g. often in excess of 50%.
• The AiLA allocation strategies are designed based on very different concepts and have therefore historically shown a low correlation to the performance of common CTA strategies.
• In this study we take the opposite approach, where the goal is to force the AiLA strategy to become positively correlated with a CTA strategy, using two difference methods.
• The purpose of the exercise is to evaluate how easy or difficult it is to force the AiLA strategy to pick up such similarity, which in turn can help to form an impression about how likely such similarity could occur un-intentionally.

• Identify AiLA allocation models which historically indicate a bias (positive or negative) wrt a given CTA signal.
• Bias metric based on daily signal frequency, compared to un-conditional frequency of the given CTA signal in the historical period, e.g. p(mo = 1) for momentum.
• For simplicity the AiLA and CTA signals are both discretized, i.e. have values -1, 0 or 1.
• The bias observed among long and short are often indicating a consistent bias and are therefore combined into a total bias metric.
• Following this approach each AiLA allocation model can be assigned a historical bias metric wrt a given CTA signal for the same futures market, e.g. model bias with respect to WTI Momentum (MO).
• Bias value can be interpreted as, "average % increase of chance of CTA signal when AiLA signal is present".

• For a given commodity and CTA signal, results indicate that certain AiLA allocation models have a bias that persists over time.
• For example, the (top-right) plot indicates how allocation signals from a certain type of model across different commodities (same color), sometimes tend to have a positive or negative bias with respect to a given CTA signal, i.e. here momentum (MO).
• The historical bias values were used in an out-of-sample test, where top/bottom bias ranked models were selected from an in-sample estimate in order to construct two, otherwise equivalent, portfolios.
• The rolling out-of-sample correlation between the two portfolios and the MO strategy are shown in the (bottom-left) plot. A slight tendency of an overall positive (negative) correlation wrt the MO strategy is shown for the max (min) bias portfolio, however, with small correlation values typically within a 25% range.

• The second method to increase the AiLA portfolio similarity to a given CTA strategy uses the same correlation constraint approach as used in earlier studies with respect to commodity (long-only) indices, e.g. see [AiLA22].
• The vector of AiLA asset weights is in this case obtained by minimizing the angular distance from the original AiLA target vector, while also satisfying a (max) angular constraint from the given CTA strategy vector.
• In contrast to the earlier study, the CTA strategy vector varies significantly over time and has long as well as short weights.
• The plot (bottom-right) shows the performance for a AiLA portfolio comprising 27 different commodities.
• The results show the impact from an increasing correlation constraint parameter (α), together with the momentum (MO) portfolio based on the corresponding signals and commodities.

• The results indicate that the method works as intended only for low correlation constraints.
• In this range the correlation (performance) tends to increase (decrease) with an increasing constraint parameter (α).
• However, when the constraint values exceed moderate levels, e.g. 0.4, the fit does not manage to find a solution that also satisfies the constraint, which results in a breakdown of correlation as well as performance.
• This is in contrast to the earlier study, with respect to commodity indices, where the method (for reasons beyond our scope here) typically were able to also yield higher correlation values.
• This inability to force the AiLA portfolio towards high correlation with the CTA strategy also suggests that the allocation signals from the two strategies are fundamentally different.

Table 1: Sharpe Ratio (SR) and Correlation (ρ) with MO for different constraint values

Figure 6: Cumulative PnL at 10% Annual Risk for different correlation constraints