Rolling Training Window Analysis

In AiLA methodology – the data is segmented into training window or period, validation (in-sample data) period and holdout (out-of-sample data) period.
Segmentation between these segments ensures no data leakage between periods and avoids overfitting.
The purpose of the exercise is to check the robustness of our methodology by testing the hypotheses.
Longer training window translates into improvement in performance.
Performance between different training windows will be correlated.

Deploy assets/constituents corresponding to 1016 assets, based on the 1117,1218,1319 and 1420 rolling training windows.
Construct different versions of two 1016 strategies 1) Diversified Commodity and 2) Single Industry: Agriculture. Construction parameters and commodity/contract maturity will be identical between the strategies, but each strategy will only have assets trained under the same window.
The performance of the strategies will be compared on a common in-sample period from 2021 onwards.

Notation used in Analysis to delineate the training windows: AAXX, where AA refers to year-end of the training window, XX refers to year-end of the validation period.
Our incumbent assets use the 1016 training window: Models are trained until end of 2010 and validation period from 2011 till end of 2016. Holdout period is from 2017 onwards.
Models will be trained under other 4 windows: Training window will be increased by one year, while keeping the validation period duration the same (6 years), while reducing the holdout period by one year.
1117: trained until end 2011 and validation period from 2012 till end 2017
1420: trained until end 2014 and validation period from 2015 till end 2020
Similarly for 1218 and 1319

Comparison 1

Commodities underlying the strategy: Agri, Base and Precious Metals, Oils: Crude and Refined products and Natural Gas.
The cumulative returns are compared across the common out-of-sample period from Jan-2021 till April-2025.
Strategy based on the 1420 cohort consistently performs better than the other cohort strategies.
Strategy based on the 1117 cohort cumulative performance is directionally similar to all the other cohorts, but underperforms the others, 1218 and 1319 have a similar performance.
1016 strategy while directionally similar, intersects the other return graphs, and also has a flatter performance from 2024.
All the cohorts have a high correlation between their returns, with 1016 having a relatively smaller correlation.

Daily return correlation

A closer look at the grouping of the returns across the rolling training windows.
1016 has a lower correlation at a monthly level compared to the daily and weekly returns.
The monthly returns correlations increase as we move from the shorter training window cohort to the longer window cohorts.

Commodities underlying the strategy: Grains, Oilseeds and Cash crops.
The cumulative returns are compared across the common out-of-sample period from Jan-2021 till April-2025.
The 1016 strategy performs well occasionally outperforming the other strategies from mid 2023.
Among the other cohorts, the 1420 performs marginally better than the others till 2023, while falling behind 1016.
1218 underperforms all the other strategies.
All the cohorts have a high correlation between their returns.

Table 2: Daily return correlation

A closer look at the grouping of the returns across the rolling training windows.
1016 has a marginally lower correlation at a monthly level compared to the daily and weekly returns.
Intermediate training windows 1117-1218, 1218-1320 have a slightly higher correlation of returns.

A key point of interest is to understand the impact of change in training window upon the ultimate performance of a strategy and to compare the performance of strategies based on differing rolling periods.
In this analysis of the two strategies, in general we see a small but discernable improvement in performance, as the training window is increased. The 1420 strategy performs marginally better than the other strategies over the period in both the cases considered. However, there are exceptions as seen in the Agri strategy comparison where 1218 underperforms.
The daily returns of all the cohorts show a high correlation between the returns and the cumulative performance of returns over the period also follow a common directional trend.
The performance of the 1016 strategy, with the shortest training window, however does not drastically differ from the other strategies and outperforms the other strategies in some time intervals.
Overall, there is a general favorable relationship between increasing training window and performance. A more definitive conclusion is that the underlying AiLA training methodology, is fundamentally consistent and therefore robust.