Frequent regime shifts in financial markets demand forecasting models that canswiftly adapt under uncertain and rapidly changing conditions. We proposeDTAM (Dynamic Task-Adaptive Meta-learning for Nonstationary DistributionShifts), whose core innovation is a detection-gated meta-objective: whenever astatistical detector flags a shift, DTAM treats this as a new ”task,” embedding thedetection signal directly into the loss function for bi-level updates. Unlike standardModel-Agnostic Meta-Learning (MAML), which presupposes stable task bound-aries and merely retrains upon external triggers, DTAM ensures that detectionoutcomes—both correct and erroneous—actively reshape the meta-parameter tra-jectory. We further provide the first theoretical guarantees of O(1/√N ) conver-gence under noisy regime detection in a nonstationary setting, along with boundedupdates under abrupt distribution shifts and high-probability consistency for thedetection process. Empirical results on benchmark equity datasets confirm thatthis detection-driven meta-learning accelerates post-shift recovery and reducesoverfitting, particularly with a Bayesian extension. DTAM thus offers a robustsolution for nonstationary financial time series.