Real return-predicting regressions in liquid markets have R-squared values around 0.01 or lower, and those can still be very profitable, because a tiny predictable component applied thousands of times compounds. An R-squared of 0.85 on next-day returns means data-leakage, not genius.
R-squared also rises mechanically when you add inputs, regardless of whether they help. Adjusted R-squared penalises the extra degrees-of-freedom and is the right one to compare models of different sizes.
The number says nothing about whether a relationship is causal or stable. Two unrelated trending series can produce an R-squared above 0.9, which is the spurious regression trap that stationarity exists to prevent.
Related: ols-regression, residuals, overfitting, stationarity