In today’s fast-paced and interconnected financial markets, the use of machine learning (ML) has become a game-changer in the realm of algorithmic trading. However, designing a consistently profitable algorithmic trading system (ATS) is challenging because of the dynamic and stochastic nature of the stock market. Often, stock markets undergo phase transitions where a market may suddenly change from a bullish trend to a bearish trend or vice versa. Most ML models fail to capture these phase changes. Consequently, inferior performance is observed from ATS during phase transitions. Moreover, the recent market data is usually small in quantity and often not sufficient to train an ATS model. We propose Adabot, an ensemble of phase-specific few-shot learners that can adapt to the changing market dynamics. Our models exploit synthetic data alongside real market data to train the ensemble. Adabot can adapt to a completely different market without any redesigning or training with extensive historical data thus reducing the deployment cycle, and can be used in markets that do not have sufficient historical data. When tested in four diverse markets, Adabot generated profits that were 30 to 90% greater than the respective benchmark returns over the test period. At the same time, Adabot significantly reduced the overall risk and did not degrade even after price shocks.