Forecasting stock prices and assessing financial risks are two intertwined challenges that remain central to modern quantitative finance. This paper presents \textbf{FIN-MIND}, a novel multi-dimensional temporal convolutional framework that jointly predicts stock prices and financial risk measures through decoupled attention mechanisms. Unlike unified-attention architectures that suffer from feature interference, FIN-MIND introduces two independent attention streams dedicated to price and risk modeling, thereby disentangling short-term momentum dynamics from long-term volatility structures. The model enforces strict temporal causality and employs a leakage-aware rolling evaluation protocol to ensure reliable out-of-sample performance. Extensive experiments on large-cap equities (AAPL and TSLA) from 2015 to 2024 demonstrate consistent improvements over classical econometric, recurrent, and transformer-based baselines. FIN-MIND achieves superior accuracy in both price forecasting and risk estimation, with significant gains across multiple random seeds. Beyond empirical performance, this framework provides a generalizable foundation for interpretable, causality-preserving financial forecasting and offers new insights into multi-task learning under non-stationary market conditions.