This paper introduces Spatio-Temporal Momentum (STMOM), a novel framework that jointly learns time-series and cross-sectional momentum strategies using multitask learning. Unlike traditional approaches that treat these strategies separately, STMOM models the trading signals for all assets simultaneously by feeding a spatio-temporal tensor of momentum features (volatility-normalized returns and MACD) into a neural network. The network is trained end-to-end to maximize the portfolio Sharpe ratio, directly outputting positions for each asset without requiring manual ranking or decile allocation. The authors evaluate several architectures—single-layer perceptron (SLP), multilayer perceptron (MLP), convolutional neural network (CNN), and long short-term memory (LSTM)—on portfolios of 46 US equities and 12 equity index futures. Backtesting from 1990–2022 (equities) and 2003–2020 (futures) shows that the simplest SLP with L1 regularization consistently outperforms more complex models and classical benchmarks (TSMOM, MACD, CSMOM, and deep momentum networks). After portfolio-level volatility scaling, the SLP achieves Sharpe ratios of 2.61 for equities and 1.73 for futures, significantly exceeding the next best method. The paper also demonstrates that STMOM retains its edge under high transaction costs (5–10 basis points) when combined with turnover regularization. The authors attribute the success of the simple SLP to the low signal-to-noise ratio of financial data and limited training samples, which cause complex models to overfit. Overall, STMOM provides a unified, data-driven approach to momentum investing that leverages cross-asset information and directly optimizes risk-adjusted returns.