Learning smooth graphs with sparse temporal variations to explore long-term financial trends

The return of inflation raises the issues of assessing cross-dependencies in the interest rates of long-term government bonds. Learning cross-dependencies directly from data is framed here as a graph learning problem that requires to address the issues of bond rates heterogeneous and nonstationary evolutions, with sharp changes along time and across countries, and of managing missing samples. As a first contribution, the present work devises a data driven time-dependent graph for bonds markets, specifically based on risk premia. As a second contribution, it shows the relevance of such constructions when applied to a broad database of 29 countries over 6 decades (1960-2020), that includes the high inflation episode of the 1970s.