Background

• AILA strategies are constructed upon assets which turn use allocation signals from dedicated classification models following a rolling multi-model (RMM) approach described below:
• From 2017-2021: Allocation signals are generated by 1016.
• From 2022, every year thereafter a new window is stacked to the existing windows combination with the signal contribution from each window considered in an equal weighted manner.
• The objective behind the approach is to maximise the utilization of finite training data while avoiding data leakage and overfitting.
• Training, deploying and operating each model in a training window involves additional investment, and therefore, necessary to assess if its beneficial.

Objective

The purpose of the exercise is to evaluate:

• Multi window approach versus single window in terms of overall portfolio performance.
• Robustness of the underlying training methodology based on consistency in overall returns upon combination.

Approach

• Diversified portfolios covering the same contracts across Agriculture - Grains, Livestock, Oilseeds, and Softs, Metals - Base and Precious, Oil - Crude and Refined Products, and Natural Gas, were constructed to simulate the gradated addition of new T-V window (Window) models each year.
• As in our portfolio construction methodology new Windows are stacked in from 2022 onwards, the analysis will consider the performance of the portfolios from that year, as the basis for evaluation.
• Accordingly, the 5 portfolios have been constructed to follow the RMM addition logic as below:
  • W0: Only include 1016 also referred to as the single Window portfolio.
  • W1: Includes 1016 and 1117 - with 50% weightage to each Window from 2022 onwards.
  • W2: Follows W1 in 2022 and 1218 added to stack with 33.33% weightage to each Window from 2023 onwards.
  • W3: Follows W2 till 2023 and 1319 added to the stack from 2024 with 25% weightage to each Window from 2024 onwards.
  • W4: Follows W3 till 2024 and 1420 added to the stack from 2025 with 20% weightage to each Window from 2024 onwards.
  • W5: Follows W4 till 2025 and 1521 added to the stack from 2025 with 16.66% weightage to each Window from 2024 onwards.

Performance Measure

• Performance of these portfolios will be measured by the cumulative returns and Sharpe ratio of the portfolios.
• It must be noted that these measures do not evaluate the performance of the individual Windows but are an assessment of the rolling methodology of stacking Windows in a gradated manner annually.
• We will also breakdown the returns at the commodity level to have a detailed view of the rolling performance for each constituent.

• All the portfolios track each other closely on the basis of overall cumulative returns and Sharpe ratios:
  • W0: 108.36% / Overall Sharpe Ratio: 3.09
  • W1: 105.69% / Overall Sharpe Ratio: 2.76
  • W2: 106.54% / Overall Sharpe Ratio: 2.73
  • W3: 104.89% / Overall Sharpe Ratio: 2.68
  • W4: 106.43% / Overall Sharpe Ratio: 2.70
  • W5: 107.28% / Overall Sharpe Ratio: 2.71
• W0, based on single Window 1016, marginally outperforms the other Window portfolios in the overall period, with W5 having all Windows coming second.
• As is expected the returns in the period 2022 to 2023 the performance are aligned followed by the period from 2023-2024 mid where W0 outperforms.
• A closer look shows that from mid 2024, W1-W5 show a sharper rate of change in cumulative returns lowering the gap with the W0 portfolio, implying a better performance from all the multi Window portfolios.
• W0 has the higher Sharpe ratio of 3.09 which sharply drops as the combination of Windows in the portfolios increases but stabilizes around 2.70.
• The daily returns have a high degree of correlation with each other, with the multi Windows being very closely correlated.

• The constructed portfolio include a total of 25 commodities. A breakdown of the analysis by commodity-wise returns was done to get a more detailed picture of the contributions.
• The two histograms presented here represent the frequency of the W0-W5 performance leader (Rank 1) based on cumulative returns and Sharpe ratio among the 25 commodities.
• While W1 at an overall portfolio level had the lowest cumulative return, ranks #1 in 6 commodities across both cumulative returns and Sharpe ratio.
• W5 while ranking #1 in 8 commodities in based on cumulative return, however ranks #1 in only 4 commodities on the basis of Sharpe ratio.
• As seen from the individual charts (See Appendix) – for each commodity, while the trend and direction of performance is similar there is a higher degree of variation in performance between W0 and W1-W5.
• The variation in Sharpe Ratio (See Appendix) between W0 and W1-W5 is also higher.
• The variation may be explained as the portfolio construction parameters, (Capital, Optimization targets, sector-asset caps etc) though identical to all 5 portfolios, may amplify variations at an individual commodity level while converging the performance at an overall level.

• A fundamental area of interest in the analysis, is to examine if a multi-Window approach to aggregating allocation signals, is better than a single Window, measured in terms of overall performance.
• In this analysis we see that a single Window portfolio – W0 outperforms the other Window combination.
• A closer looks reveals some interesting patterns - While there does seem to be a drop in performance as Windows are stacked in the first 2 years, suggesting a return degradation with respect to W0. From 2024 onwards we observe the performances catch up and converge suggesting that the Windows are return additive in this period.
• Examining the performance at the individual commodity level – we see that while single Window returns have the highest frequency of outperformance, the other Window combinations (W1 and W5) also perform credibly for certain commodities. However, the variation in performance between the Window combinations is higher.
• While we cannot conclusively determine that there is a positive relationship between adding more Windows, but the rolling windows do not degrade the absolute returns or the quality of returns significantly, while adding a buffer in shoring up performance in certain periods when the single Window portfolio falters.
• A more definitive conclusion, based on the performance of W0 and the high correlation between the performances of the W0-W5, is that the underlying AILA training methodology is highly consistent and therefore robust.