The team undertook a major research project into rotation between different factors. As a result of this research, they managed to introduce gradual improvements into their investment process, and to make it more robust.
The Jupiter systematic Team
Jupiter Systematic Equities team, headed by Amadeo Alentorn, and including Tarun Inani, James Murray, Yuangao Liu, Sean Storey and Matus Mrazik (from left to right).
Factor rotation has long been an integral part of our investment process. Over the years we have been able to demonstrably add value through our dynamic weighting scheme by overweighing factors in environments we have forecasted to be favourable, and vice versa. Interestingly, successful factor rotation has not yet been generally accepted by either investment professionals or the academic community. One can make an analogy with market efficiency, where many managers either do not believe in active management or in some cases hide behind small bets, not deviating enough from their benchmarks. At Jupiter we have a strong commitment to active fund management, which in the case of the Systematic Equities team is demonstrated by our proprietary stock selection criteria – alpha factors – but perhaps more notably by our dynamic weighting scheme, which has the ability to significantly depart from a static factor allocation. Despite our success with dynamic rotation, we do recognise challenges in timing factors, which has resulted in our continuous research and gradual improvements in this area.
Model upgrade
Our recent model upgrade is the culmination of a research project focusing on better identification of relationships between our market sentiment and risk environment variables and our factor set. Successful factor timing relies on first identifying a predictive relationship and the subsequent translation of this information into factor and stock level forecasts and trades. It is well understood that one does not need to forecast with perfect accuracy to make a success of a timing model. In practice, the best that we can hope for is to classify and predict that a factor moves correctly “more often than not” and thus get closer to the unattainable goal of a perfect forecast. The first research objective is therefore to increase the hit-ratio of correct factor forecasts. Optimal factor bet position sizing is also crucial in order to feed the factor timing skill into the portfolio optimisation process and thus into the final stock level forecasts. Having more certainty about a predictive relationship implies more conviction, larger deviation from a static allocation and thus larger expected factor rotation size. Analogously to tracking error in portfolio construction, larger deviations from static allocation will result in more active risk from factor rotation. If the factor forecasts carry predictive information (skill) this will translate into alpha.

Specifically, this research focused on estimating certainty around the predictive relationships. We make larger forecasts in factors where the certainty is higher, and vice versa. Moreover, we have also increased the overall rotation ‘amplitude’ to reflect our overall higher conviction in the rotation model. We have been able to make this improvement due to the conviction weighting of the individual factors resulting in overall higher confidence in the model, allowing us to take higher active risk from factor rotation. From a portfolio construction perspective, this model improvement reduces noise and unnecessary model turnover, and focuses alpha where pockets of predictability can be found. This is evidenced by our simulations showing increased absolute and risk adjusted return.

To arrive at the conviction metrics, we have looked for consistency across time and across regions. Validation of statistical hypotheses, as well as economic priors in these two dimensions, is a helpful way to maximise robustness and attempts to maximise the potential for the model to generalise well in the future. Our original model was designed more than ten years ago and although performing well out-of-sample, we have now had an additional ten years of data to cross-validate the model. The new data has confirmed many strong predictive relationships, but some showed weaker linkage evidence, leading to a smaller likelihood of predicting factor moves. It is worth noting that although there is significant correlation in the market sentiment and risk environment variables (mostly due to interconnectedness and globalisation of financial markets), at times we observe significant differences between regions. It is therefore important to make sure that the predictive relationships hold for individual regions as well as in aggregate.
Increasing robustness
We have also improved the robustness of how market sentiment and the risk environment are linked with subsequent factor returns by avoiding overreliance on a small number of points in history. For example, if a high (or low) risk environment has historically been followed by some extreme factor moves, this could imply that periods with similar level of risk environment could be expected to be followed by similarly extreme outcomes. However, although we look for similarity across time, there also needs to be recognition that certain periods might be unlikely to repeat with the same magnitude. We have explored several ways to address this problem and arrived at a return transformation in the linkage estimation that treats returns more like ranks in history. This way the model links, for example, high market sentiment with high subsequent returns but not necessarily with an absolute quantity. We have demonstrated that this approach further robustifies our approach to factor rotation.

These changes together not only improve our current model, but they have also been made with further research in mind. We hope to be able to add new conditioning variables, where identifying new relationships, and their significance, as well managing model noise and turnover will be crucial.