Highlights:

• kNN (k Nearest Neighbours) is a machine learning  classification algorithm. For our TAA use case this algorithm matches today’s macro environment to its  closest historical look‑alikes and uses the market  moves that followed those periods to derive forecasts.  
• Two kNN models assessing the relative attractiveness  of Equities vs Government Bonds are already embedded in our Research TAA process. Over the past two  years, the US model achieved 83% accuracy (see  chart), and the euro‑area model 63%.  
• By “trading” Equities against Government Bonds,  these two models already target one of the most rewarding levers in classical TAA. Other binary classification tasks such as High Yield Credit vs Investment  Grade Credit are also worth scrutiny. However, simply combining isolated pairwise comparisons lacks the ability to  capture interactions across pairs adequately. In a multi-asset class universe comprising Equities (EQ), High Yield Credit  (HY), Investment Grade Credit (IG), Government Bonds (GV), and Cash (CS), jointly evaluating all five asset classes  would have lifted the TAA potential by more than 25% compared with just combining the two pairs EQ vs GV and HY vs IG.  
• Extending the algorithm beyond a two‑asset class framework requires overcoming additional methodological challenges. In the two‑asset class case there is only one degree of freedom: classifying EQ as the outperformer automatically implies GV as the underperformer, and vice versa. In the five‑asset class case, we must determine a joint ranking  across all asset classes while (i) using a single, shared neighbourhood for each asset class and (ii) enforcing con-  sistency of the forecasted ranks.  
• Our five‑asset class kNN model was trained and tested on monthly US macro data from the FRED‑MD. In a true out‑of‑sample check covering 01/2022–10/2025, the signal‑based TAA delivered +7.2 bps per month of added value, with a 75% success rate over the past two years. 

Download the full publication below