Quantum vs. Classical Machine Learning: A Benchmark Study for Financial Prediction

TL;DR

The paper tackles the question of whether current generation quantum machine learning can offer practical advantages for financial forecasting. It introduces a reproducible benchmark that pairs architecture matched quantum and classical models across three tasks in two markets: directional prediction, live trading, and realized volatility forecasting. Across tasks, quantum methods show gains when circuit design and data embeddings align with the task structure, with QNNs improving recall in high dimensional feature spaces, QLSTMs delivering regime-dependent risk-adjusted improvements, and QSVR providing competitive volatility forecasts relative to classical kernels. The work provides actionable guidance on when quantum methods add value and offers a standardized framework to benchmark future QML advances in finance.

Abstract

In this paper, we present a reproducible benchmarking framework that systematically compares QML models with architecture-matched classical counterparts across three financial tasks: (i) directional return prediction on U.S. and Turkish equities, (ii) live-trading simulation with Quantum LSTMs versus classical LSTMs on the S\&P 500, and (iii) realized volatility forecasting using Quantum Support Vector Regression. By standardizing data splits, features, and evaluation metrics, our study provides a fair assessment of when current-generation QML models can match or exceed classical methods. Our results reveal that quantum approaches show performance gains when data structure and circuit design are well aligned. In directional classification, hybrid quantum neural networks surpass the parameter-matched ANN by \textbf{+3.8 AUC} and \textbf{+3.4 accuracy points} on \texttt{AAPL} stock and by \textbf{+4.9 AUC} and \textbf{+3.6 accuracy points} on Turkish stock \texttt{KCHOL}. In live trading, the QLSTM achieves higher risk-adjusted returns in \textbf{two of four} S\&P~500 regimes. For volatility forecasting, an angle-encoded QSVR attains the \textbf{lowest QLIKE} on \texttt{KCHOL} and remains within $\sim$0.02-0.04 QLIKE of the best classical kernels on \texttt{S\&P~500} and \texttt{AAPL}. Our benchmarking framework clearly identifies the scenarios where current QML architectures offer tangible improvements and where established classical methods continue to dominate.

Figures (9)

• Figure 1: Overview of financial decision system, motivations and contributions of our work.
• Figure 2: A detailed overview of our methodology. We assemble a Portfolio Universe that spans U.S and Turkey Markets and then fetch daily prices with the Yahoo Finance API. Each branch applies task‑specific feature engineering. Features are subsequently scaled and split by cross‑validation scheme before being fed into paired quantum and classical models: QNN vs ANN, QLSTM vs LSTM, and QSVR vs SVR. Downstream arrows indicate the learning target for each study and the evaluation metrics used to compare models.
• Figure 3: We use 3 feature regimes in our first study. Adj Close refers to stock returns computed from adjusted‑close prices, which account for splits and dividends. Lag indicates the previous day's return of the stock. (A) For Turkish equities, a 3-D feature set of technical indicators is constructed from adjusted close prices. (B) For S&P 500 index, a 7-D feature set is derived from cross-asset relations. (C) For selected US equities, a 64-D feature set includes the past 8 day returns of the equity itself as well as the 7 day returns of eight major indices.
• Figure 4: Walk-forward cross-validation splits used across five folds for directional classification. Each fold uses an expanding training window followed by a fixed validation window.
• Figure 5: QNN architectures explored for directional classification. A classical feature set of dimension $d$ is fed into the model. If $\texttt{Hybrid} = \texttt{True}$, the QNN optionally preprocesses these features using a classical layer that maps the $d$ inputs to $q$ outputs. The resulting $q$-dimensional feature vector is then encoded into a register of $q$ qubits via angle or amplitude encoding. A parameterised ansatz of rotation gates ($R_x$, $R_y$, $R_z$) and entangling operations is applied. After the circuit, either (1) a single-qubit readout measures one qubit and the expectation value is normalised from $[-1, +1]$ to $[0, 1]$, or (2) a multi-qubit readout measures all $q$ qubits and aggregates the results. The post-processed measurement outcome yields the final predicted label.