The study evaluates the effectiveness of metamodeling in predicting the sign of next-day returns for 30 assets between 2001 and 2024. In essence, we compared using 10-year rolling windows five single models (ARIMA, Logistic Regression, Random Forest, XGBoost and LSTM) with a metamodel designed to predict the sign using two rules – majority, if four out of five models predict the same sign, or a fallback mechanism if there is no consensus among individual models. The performance is evaluated using 3 indicators: sign accuracy, Sharpe ratio and cumulative returns. A separate analysis was carried out for the period of the global financial crisis. Results indicate that the Metamodel generally provides robustness and performance similar to, but slightly worse than that of the best single model. The situation is the same when we take in consideration the global financial crisis period. The performance of the Metamodel is better when analyzing equities from the technology sector or stock indices. Thus, metamodeling is useful as a stabilizing instrument and to reduce the model selection risk, but the benefits depend significantly on market regimes or asset classes.