Imagine your computer not only analyzing financial charts but also learning to make investment decisions on its own – faster and smarter than a human. Now add a touch of quantum physics. Sounds like science fiction? Yet, recent research shows that combining reinforcement learning, quantum-inspired neural networks, and classical financial data can provide a real edge in trading.

This is exactly the focus of a publication from National Taiwan Normal University and Wells Fargo. The researchers built a trading agent that uses quantum-enhanced neural networks to trade the USD/TWD (US Dollar/Taiwan Dollar) currency pair.

Core Idea

Let’s start with analogies:

  • Currency trading is like running a shop: you buy cheap and sell high. The problem is that prices change every second, and predicting tomorrow is extremely difficult.
  • Artificial intelligence in trading acts like a super-smart advisor who looks at charts, learns from history, and suggests: “buy now”, “sell now”, or “wait”.
  • Reinforcement learning (RL) is like dog training: the agent gets rewards for good actions (profit) and penalties for bad ones (loss). Over time, it learns better strategies.
  • Quantum neural networks (QNN, QLSTM) act like additional senses – they can capture subtle patterns in noisy data that a classical model might miss.

The study combined two elements:

  1. QLSTM – a special network predicting short-term trends (whether the price will go up or down by at least 1.2% in the next 5 days).
  2. QA3C – a trading agent deciding whether to buy, sell, or hold, based on QLSTM predictions and its own state features.

The result? In 5 years of testing (2020–2025), the agent achieved 11.87% profit with a maximum drawdown of just 0.92%. In practice, it earned steadily with very small risks – outperforming many currency ETFs.

👉 In practice: such a system could be used for automated investing in currencies, stocks, or cryptocurrencies, minimizing the risk of sudden large losses.

Formalization and Technical Details

1. Architecture

Researchers applied a hybrid approach:

  • QLSTM (Quantum LSTM) as a feature extractor – input sequences of 4 days (open, high, low, close, MA5, MA10). Each value was encoded as quantum rotation gates $Ry(\theta)$. The output was a binary prediction (“up”/“down”), later used as features.
  • QA3C (Quantum Asynchronous Advantage Actor-Critic) – RL agent with three actions: hold, buy, sell. The state vector included 10 features (e.g., QLSTM probabilities, cash ratio, PnL, deviations from MA20/MA60). The policy network contained a classical layer + 8-qubit VQC.

2. Problem Setup

QLSTM Forecasting

Each day $t$ receives a label:

$$ y_t = \begin{cases} 1 & \text{if } \frac{y(t+5) - y(t)}{y(t)} > 0.012 \\ 0 & \text{if } \frac{y(t+5) - y(t)}{y(t)} < -0.012 \\ \text{none} & \text{otherwise} \end{cases} $$

Training: RMSprop, learning rate $5 \times 10^{-3}$, 50 epochs, weighted cross-entropy. Test accuracy ≈ 71.5%.

QA3C – Action Selection

Action space:

$$ A = {0: \text{hold},; 1: \text{buy},; 2: \text{sell all}} $$

Policy and value functions:

$$ \pi_\theta(a|s_t), \quad V_\theta(s_t) $$

Update rule with $n=30$ step rollouts:

$$ L = \delta^2 + \left(-\log \pi_\theta(a|s)\right)\delta - \beta H[\pi_\theta(\cdot|s)] $$

where $\delta = R - V_\theta(s)$ and $H$ is entropy for exploration.

3. Reward Function

The reward consisted of:

  • time cost ($-0.02$ per day),
  • entry rewards/penalties depending on trend alignment,
  • exit bonus: $ +10 + 50 * PnL%} $, for profitable sells,
  • exit penalty: $ -2 - 10 * PnL% $ for losses,
  • holding penalty: $ -5 * (PnL%)^2 $.

Clipping: $r_t \in [-15, 30]$.

4. Results

  • QA3C: 244 parameters (32 quantum, 212 classical).
  • A3C (classical): 3332 parameters.
  • Profit: QA3C – 11.87%, A3C – 11.42%.
  • Max drawdown: both 0.92%.
  • Win rate: QA3C 56.7% vs A3C 54.8%.

Conclusion

The publication shows that quantum neural networks are no longer just theory – they can already improve trading algorithms even when simulated on classical computers.

Key takeaways:

  • Hybrid approach (QLSTM + QA3C) leads to more stable and efficient trading strategies.
  • 11.87% profit in 5 years with minimal drawdown is competitive with major currency ETFs.
  • Applications extend beyond forex – similar models could support portfolio management, derivatives pricing, and risk analysis.
  • In the future, when quantum computers become available, the effects may be even stronger.

👉 That’s why it’s worth keeping an eye on quantum machine learning (QML) – the near future of investing may belong not only to AI, but to quantum-enhanced AI.

📎 References