Validation-Gated Knowledge Transfer for Heterogeneous Federated Time-Series Models

Federated learning (FL) commonly relies on homogeneous client architectures and parameter averaging, assumptions that may be unsuitable when participating institutions employ different predictive models and exhibit strongly non-independent and identically distributed (non-IID) data. We investigate this problem through institution-specific foreign-exchange quotation modelling and propose a validation-gated knowledge-transfer framework for heterogeneous federated time-series models. Three simulated institutional clients are constructed using real USD/LKR market observations together with semi-synthetic private factors representing liquidity, inventory, demand, and order-flow behaviour. Client-specific GRU, LSTM, and 1D-CNN models are selected using local validation performance, enabling architectural heterogeneity with negligible loss relative to unconstrained model selection. Rather than averaging incompatible model parameters, client predictions on a shared multi-profile anchor dataset are aggregated to form collaborative knowledge. We further introduce target-specific soft blending, in which the contribution of collaborative knowledge to midpoint-adjustment and spread predictions is independently selected using local validation data. Preliminary experiments show that homogeneous FedAvg performs substantially worse than heterogeneous local modelling. For the most challenging client, validation-selected performance-weighted blending reduced spread MAE from 24.12 to 23.24 basis points, while also reducing buying-rate and selling-rate errors by approximately 3.5%. Across the three clients, the same strategy reduced average spread MAE from approximately 13.67 to 13.37 basis points. In contrast, full distilled-model updates were rejected by the validation safeguard, indicating that unrestricted knowledge transfer can cause negative transfer under strong client heterogeneity. These preliminary findings suggest that selective prediction-level collaboration may provide a safer alternative to homogeneous parameter aggregation for heterogeneous federated learning.