Portfolio Optimization with TD3 Algorithm Using Technical Indicators

We explore the application of the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm to portfolio optimization. Financial markets present high-dimensional and non-stationary environments where traditional reinforcement learning (RL) methods often suffer from value overestimation and instability. Unlike other approaches that use OHLCV (Open, High, Low, Close, Volume) data, we evaluate our approach using only a diverse set of normalized technical indicators. Such indicators are derived from OHLCV data; however, they are widely used by investors who rely on technical analysis. Our experimental results demonstrate that the agent achieves stable convergence. We further analyze the policy’s behavior using the Herfindahl–Hirschman Index (HHI) and portfolio turnover, revealing a shift toward more concentrated capital allocation during the inference phase.