Q-PhotoMarket: A Design Space Exploration Framework for Photonic Hybrid Quantum Neural Networks in Financial Market Prediction

AI summary

Overview

Research area: Quantum machine learning — specifically photonic (linear-optical) hybrid quantum neural networks applied to financial time-series prediction, with a design space exploration (DSE) methodology.

Technical level: Intermediate. The paper assumes familiarity with hybrid quantum-classical models, photonic circuit primitives (beamsplitters, phase shifters, interferometers), and standard ML evaluation metrics, but its central argument — that architecture choice matters — is accessible to a general ML reader.

Scope (one sentence): The paper systematically explores roughly 5,000+ photonic HQNN configurations spanning input photon states, circuit architectures, entangling models, and measurement strategies across U.S., Indian, and cryptocurrency markets, and reports which design dimensions actually move predictive performance.

What This Paper Is About

Photonic quantum computers offer room-temperature operation and native linear-optical circuits, making them attractive for hybrid quantum machine learning, but existing photonic QML studies for finance typically evaluate only a single hand-picked architecture. That leaves open the question of whether reported results reflect the capability of photonic HQNNs or simply an unlucky (or lucky) circuit design. The paper builds Q-PhotoMarket, a framework that maps out the photonic design space and evaluates thousands of valid configurations under one unified protocol, so that architectural trends can be separated from noise.

Key Contributions

1. A DSE framework for photonic HQNNs. Q-PhotoMarket organizes the photonic design space into four independent dimensions — input photon state, circuit architecture, entangling model, and measurement strategy — with the computation space determined by the measurement choice.

2. Large-scale empirical evaluation. Thousands of valid HQNN configurations (approximately 600 per market/threshold pair, roughly 5,400 total runs) are evaluated across three markets and three return thresholds under a leakage-safe, expanding-window walk-forward protocol.

3. Hybrid search strategy. Exhaustive grid screening is combined with Bayesian optimization (Optuna's TPE sampler) to reduce the cost of searching the high-dimensional space, including a warm-starting stage that seeds local refinement from the best grid results.

4. Calibration and collapse diagnostics. Threshold calibration plus a prediction-collapse detector make evaluation reliable under the severe class imbalance that appears at higher return thresholds, and the paper provides empirical insights into which photonic design choices actually matter.

Main Findings

• Design choices matter enormously. The distribution of AUC across configurations is broad; the median stays near the random-ranking reference (0.500) while the best configurations reach notably higher, showing that manually picking one architecture can mislead about photonic HQNN capability.

• Circuit topology is the dominant factor. The Clements Mesh consistently achieves the strongest average performance, particularly for S&P 500 and cryptocurrency; Ring NN is competitive; the Phase-only architecture is generally weakest.

• Measurement strategy is the other major lever. Fock-basis probabilities perform strongly on S&P 500 and NIFTY 50, while correlation-based measurements become increasingly competitive on crypto — so the optimal measurement representation is application-dependent.

• Input photon state has modest influence. The four input states (Alternating, Bunched, Pair-bunched, Thermal-like) show similar GPS and AUC distributions across all three markets.

• Entangling model matters less than topology. Bell and MZI distributions overlap substantially, with MZI showing a slight advantage on NIFTY 50.

• Higher return thresholds improve ranking but worsen imbalance. Mean AUC rises from 0.500 at τ=0 to 0.528 at τ=1% for S&P 500, from 0.499 to 0.517 for NIFTY 50, and from 0.500 to 0.510 for crypto — because economically significant moves are more separable even though the positive class drops to 20–34% prevalence.

• Best configurations per market. S&P 500: Simple circuit + Bell entangling + Fock probabilities, mean AUC 0.528 (± 0.019). NIFTY 50: Clements Mesh + MZI + Fock measurements, mean AUC 0.514 (± 0.028). The paper does not state a single headline best configuration for crypto in the text; its top-10 table for crypto is dominated by Clements Mesh, MZI/Bell, and expectation/correlation measurements.

• Longer training helps shortlisted models. Phase 2 training (20 epochs versus 8 in Phase 1) consistently improves median fold-wise AUC while fold variance stays comparable.

• Class weighting is market-dependent. S&P 500 and NIFTY 50 perform best with class-weighting=balanced, whereas crypto benefits from class-weighting=none.

• Prediction collapse is a real failure mode. At τ>0, predicted probabilities often stay below 0.5, sending all samples to the majority class — AUC remains meaningful but F1, MCC, and balanced accuracy collapse (F1=0, MCC=0, BalAcc=0.5). Calibration recovers meaningful positive prediction rates.

• Competitive but modest versus classical baselines. The best photonic HQNN consistently beats the majority-class baseline and is competitive with logistic regression and a shallow MLP; the gap widens at higher thresholds where simple classifiers struggle with minority positives.

Methodology in Plain English

The researchers treat architecture selection itself as the experiment. They fix the input to 16 price-derived features per asset, which with the extra mode gives 17 photonic modes, and hold training constant across every configuration: Adam at a learning rate of 3×10⁻⁴, binary cross-entropy loss, batch size 128, 100 measurement shots, seed 42, 8 epochs in the screening stage and 20 in the confirmation stage.

They then enumerate the cross-product of four design dimensions — 4 input photon states, 5 circuit topologies, 2 entangling primitives, and 5 measurement strategies — and discard combos that don't make physical sense (for example, non-binary input states paired with unbunched or dual-rail probability measurements). That leaves roughly 600 valid configurations per market-and-threshold, times 3 markets and 3 thresholds, for about 5,400 runs; about 1,080 configurations are screened per market and about 600 re-evaluated with multi-fold validation.

To combat look-ahead bias, labels (whether the next-day return exceeds a threshold τ of 0, 0.5%, or 1.0%) are constructed only after feature engineering, and evaluation uses an expanding-window walk-forward scheme with 5 folds and a minimum training fraction of 0.5. Because higher thresholds make positives rare, they apply inverse-frequency class weighting and calibrate the decision threshold on the validation fold by maximizing MCC, F1, or balanced accuracy, recording the positive prediction rate and flagging any run outside the 0.02–0.98 band as a collapse.

Finally, they layer Bayesian optimization on top using Optuna's TPE sampler — 150 trials per shard, 30 random startup trials, a 300-second per-trial timeout, a composite objective weighting AUC 0.6, F1 0.3, and GPS 0.1 with a 0.15 penalty for collapse, plus a resource guardrail and a median pruner with 5 warmup epochs — and warm-start later studies with earlier grid results. Everything runs in simulation via the Merlin photonic framework with PyTorch autodifferentiation, so no parameter-shift estimation is needed.

Why This Matters

For research: The paper reframes photonic QML evaluation — the headline result is not a single architecture but the demonstration that circuit topology and measurement strategy produce larger performance variation than input-state or entangling-primitive choices. That suggests future photonic QML work should invest in architecture search rather than isolated circuit tuning, and it supplies a reproducible benchmark for doing so.

Real-world applications:

• Equity signal generation for U.S. (S&P 500) and Indian (NIFTY 50) markets, where the framework reports which circuit/measurement combinations generalize across thresholds.
• Cryptocurrency trading models, where correlation-based measurements unexpectedly outperform the probability-based choices that win on equities.
• Regime-aware risk tooling, since the paper explicitly handles imbalanced, high-threshold "economically significant move" prediction rather than plain up/down direction.
• Benchmarking infrastructure for groups deciding whether a photonic backend is worth pursuing relative to logistic regression or a shallow MLP.

Industry relevance: The use of a class-weighting setting that differs by market, plus threshold calibration and collapse detection, mirrors practical deployment concerns in quantitative finance — a model that always predicts "no significant move" looks fine on AUC but is useless. The explicit comparison against cheap classical baselines sets a realistic bar for photonic hardware claims.

Future Directions

1. Scaling beyond simulation. All experiments use differentiable photonic simulation; extending to hardware-aware optimization and execution on real photonic devices is the stated next step.

2. Larger and broader design spaces. The explored space is limited to the investigated topologies, measurement strategies, and encodings — larger photonic architectures and additional encodings remain open.

3. Automated architecture search. The authors point toward automated search techniques to further cut the cost of exploring high-dimensional photonic design spaces, building on the TPE-based stage here.

4. Application transfer. Applying Q-PhotoMarket beyond daily market prediction to other real-world QML domains, and examining settings where signals are less weak than in financial time series.

Target Audience

Researchers and graduate students working on quantum machine learning who want to know which photonic design decisions are worth their time; quantum-hardware and compiler engineers evaluating linear-optical platforms for ML workloads; and quantitatively minded finance practitioners who need to judge whether photonic QML offers anything beyond logistic regression under rigorous walk-forward evaluation. Readers seeking a definitive "quantum advantage" claim will not find one here — the framing is deliberately about characterizing a design space rather than beating classical baselines.

Authors’ abstract

Photonic quantum computing has recently emerged as a promising platform for hybrid quantum machine learning due to its native realization of linear-optical circuits and the computational complexity of boson sampling. However, despite growing interest in quantum methods for finance, the influence of photonic circuit design choices on predictive performance remains largely unexplored. Existing studies typically evaluate a single architecture, leaving the broader photonic design space unexamined. In this work, we present Q-PhotoMarket, a systematic design space exploration (DSE) framework for photonic hybrid quantum neural networks (HQNNs) applied to financial market prediction. We explore over 5,000 valid photonic configurations spanning input photon states, circuit architectures, entangling models, and measurement strategies across their compatible computation spaces, for U.S., Indian, and cryptocurrency markets. To improve search efficiency, the exhaustive exploration is complemented with Bayesian optimization. We further incorporate threshold calibration and prediction-collapse diagnostics to enable reliable evaluation under increasingly imbalanced return thresholds. Experimental results show that systematic exploration of more than 5,000 photonic HQNN configurations reveals consistent architectural patterns across financial markets, identifies robust high-performing designs, and demonstrates competitive performance relative to classical machine learning baselines.