LightQuant: Quantitative Finance & Quantum Sensing

• LightQuant is a dual-meaning concept that serves as a lightweight, user-friendly framework for Chinese multimodal stock analysis and integrated backtesting.
• It features a modular architecture with distinct data, model, and evaluation layers, enabling rapid prototyping and robust financial prediction.
• In broader usage, LightQuant refers to quantum light sensing and imaging methods leveraging induced coherence, correlation structures, and low-photon technologies.

LightQuant most specifically denotes the lightweight, user-friendly quantitative investment and backtesting framework introduced alongside the CSMD multimodal dataset for Chinese stock analysis. In that explicit sense, it is a practical research platform for financial data analysis, feature engineering, model training, prediction evaluation, and backtesting/simulation, designed to simplify experimentation relative to more complex open-source quant platforms such as Qlib (Liu et al., 3 Nov 2025). At the same time, the supplied literature repeatedly uses “LightQuant” as a contextual label for quantum-light sensing, imaging, and ranging systems rather than as a single formal platform name. This suggests a broader, informal umbrella usage in which “LightQuant” refers to quantitative information extraction from light, including quantum illumination, induced-coherence ranging, low-photon imaging, and phase-sensitive optical metrology (Zhao et al., 2021).

1. Definition and disambiguation

In the explicit arXiv record, LightQuant is a framework for Chinese multimodal stock analysis, not a forecasting model in itself. It is introduced with CSMD as a lightweight, modular environment for end-to-end experimentation on aligned price and financial-news data, with particular emphasis on accessibility for researchers and practitioners who want rapid prototyping without the tooling overhead of larger systems (Liu et al., 3 Nov 2025).

A second usage appears across the supplied photonics literature. There, “LightQuant” functions as a descriptive tag for light-based quantitative sensing and imaging problems, including quantum target detection, LiDAR, low-light imaging, nonclassicality measurement, and quantitative phase imaging. This suggests a naming overlap rather than a single cross-domain technical system (Zhuang et al., 2017).

Usage	Domain	Representative source
LightQuant	Chinese multimodal stock analysis and backtesting	(Liu et al., 3 Nov 2025)
“LightQuant”-style usage	Quantum sensing, LiDAR, imaging, and optical quantification	(Zhao et al., 2021)

A common source of confusion is the proximity of the name to LightQANet, which is a separate low-light image enhancement framework based on quantized and adaptive feature learning, with the Light Quantization Module and Light-Aware Prompt Module as its central components (Wu et al., 16 Oct 2025). LightQANet is not the same system as LightQuant.

2. Origins and motivation in Chinese stock analysis

The financial LightQuant framework was designed to address a practical gap in Chinese stock research. The paper identifies several motivations: reducing tooling complexity, supporting multimodal Chinese stock research, enabling rapid prototyping, and providing a framework that works directly with the curated CSMD datasets (Liu et al., 3 Nov 2025). This positioning is explicit: LightQuant is presented as a simplified alternative to more complex open-source quant platforms such as Qlib, which the paper describes as powerful and end-to-end but potentially too complex for newcomers.

The framework is tied to the observation that many available benchmark datasets and tools are geared toward the U.S. market and English-language financial text. LightQuant is therefore coupled to CSMD as infrastructure for a different problem setting: Chinese-language financial news, Chinese market structure, and stock analysis under multimodal alignment of news and price data (Liu et al., 3 Nov 2025). Its intended function is not merely data storage or benchmarking; it is the operational layer through which CSMD becomes usable for stock movement prediction and trading/backtesting validation.

The paper also frames LightQuant as a way to let researchers focus on modeling and financial insight rather than data plumbing, baseline implementation, and backtesting infrastructure. In this sense, its main contribution is engineering and usability rather than the proposal of a new predictive algorithm (Liu et al., 3 Nov 2025).

3. Architecture, workflow, and supported research tasks

LightQuant follows a modular three-layer architecture consisting of a Data layer, a Model layer, and an Evaluation layer (Liu et al., 3 Nov 2025). The Data layer provides a unified interface for data extraction, data processing, storage, and heterogeneous source handling. It supports multi-dimensional market data, financial news text, feature engineering, and factor libraries. The Model layer supports development and integration of predictive models, model loading, training, and invocation, and is described as plug-and-play. The Evaluation layer provides evaluation metrics, backtesting, performance analysis, and strategy validation/refinement tools.

The workflow described in the paper is correspondingly linear. LightQuant ingests the CSMD datasets, specifically CSMD 300 for CSI 300 component stocks and CSMD 50 for SSE 50 component stocks. It then supports preprocessing and feature engineering, model training, prediction evaluation, and finally backtesting through trading simulation (Liu et al., 3 Nov 2025). The supported modalities include historical stock prices, aligned Chinese financial news text, and LLM-enhanced factor representations extracted from news.

The framework hosts both single-modal and multimodal baselines. The single-modal models listed are LSTM, BiLSTM, ALSTM, Adv-LSTM, SCINet, and DTML; the multimodal models are StockNet, HAN, and PEN (Liu et al., 3 Nov 2025). Prediction is evaluated with ACC and MCC, while backtesting is evaluated with ARR, Sharpe Ratio, MDD, and Calmar Ratio. This organization makes clear that LightQuant is an experimentation substrate spanning data preparation, modeling, and trading simulation, rather than a single inference architecture.

4. Data curation, validation, and empirical results

CSMD is the dataset substrate on which LightQuant operates, and the paper treats data quality as a central design variable. News text is collected from Securities Times, described as a reputable financial media source, while price data are collected from Baostock (Liu et al., 3 Nov 2025). The framework uses not only raw text but also LLM-extracted factors derived from carefully crafted prompts, with the aim of obtaining more interpretable and financially aligned factor representations.

The data are validated along five axes: denoising, financial sentiment expression, text density, human readability, and LLM readability (Liu et al., 3 Nov 2025). Manual evaluation is performed by five financial experts who judge coherence, relevance, and accuracy. Automatic evaluation uses MiniLM-L6-v2 for text sorting and GPT-4 scoring on coherence, information content, and topic depth. The resulting claim is that LightQuant benefits from a higher-quality input pipeline than prior datasets centered on Chinese multimodal stock analysis.

The empirical results reported in the summary indicate that CSMD 300 and CSMD 50 outperform CMIN-CN across most models, which the paper attributes to higher-quality text, better denoising, richer multimodal alignment, and LLM-enhanced factor extraction (Liu et al., 3 Nov 2025). On CSMD 300, the best ACC is reported as StockNet = 55.47, while the highest MCC in the table is DTML = 0.1478. On CSMD 50, the best ACC is reported as StockNet = 55.11, while the accompanying summary notes an inconsistency regarding the highest MCC: the table reports 0.1002 as the top value, but the text does not resolve cleanly whether that score should be assigned to PEN or HAN (Liu et al., 3 Nov 2025). In backtesting on CSMD 50, averaged across 50 stocks, StockNet achieves the best ARR = 0.1301 and CR = 3.0182, ALSTM the best Sharpe Ratio = 0.8192, and HAN the best MDD = 0.0149 (Liu et al., 3 Nov 2025).

These results position LightQuant as an evaluation environment in which dataset quality and multimodal alignment are treated as first-class determinants of downstream stock prediction and trading performance.

5. Quantum ranging and target detection in the broader “LightQuant” usage

In the broader light-based usage suggested by the supplied literature, one major cluster concerns noisy target detection and LiDAR. Quantum illumination for Rayleigh-fading targets analyzes target presence/absence discrimination when the return has Rayleigh-distributed amplitude and uniformly distributed phase. In that setting, the optical parametric amplifier receiver loses its advantage because random phase averaging destroys the phase-sensitive cross-correlation it requires, whereas the sum-frequency generation receiver still outperforms the classical benchmark. The paper’s central result is that the advantage under Rayleigh fading becomes subexponential rather than exponential, with the SFG receiver’s error probability lower than the classical system’s by roughly a logarithmic factor when $M\bar{\kappa}N_S/N_B \gg 1$ (Zhuang et al., 2017).

A second line of work formulates target detection as estimation of probe transmission. The Quantum Time-Correlation protocol uses SPDC photon-pair sources, time-resolved photon counting, Fisher information, and ROC analysis to show how phase-insensitive temporal correlations between probe and reference photons improve detection relative to classical intensity detection, especially in noisy, lossy environments (Liu et al., 2020). A related covert-ranging perspective appears in Quantum Rangefinding, which exploits the fact that one arm of a non-degenerate two-mode squeezed SPDC state is thermal when viewed alone. The transmitted probe therefore appears as a maximally mixed or thermal state, while range is recovered from coincidence-time histograms and frequency anti-correlations; the experiment uses a 750 ps bin width corresponding to about 10 cm range resolution (Frick et al., 2020).

More recent quantum LiDAR architectures shift the sensing observable itself. “Quantum Light Detection and Ranging” uses entangled photon pairs and time-resolved coincidence detection with a SwissSPAD2 time-gated single-photon avalanche diode camera. The camera gate is scanned in 18 ps steps over about 27 ns with a gate width of roughly 15 ns, giving a depth resolution on the order of 2.7 mm, and the system separates genuine returns from synchronous and asynchronous spurious signals without prior knowledge of the scene (Zhao et al., 2021). “Quantum Induced Coherence Light Detection and Ranging,” inspired by the Zou-Wang-Mandel experiment, avoids direct detection of the reflected probe photons and instead detects only local reference photons whose single-photon interference fringes encode object distance. The reported system achieves 412 μm lateral resolution, 5.1 μm ranging accuracy, and about a 20 dB enhancement in noise resilience relative to a classical-coherence approach (Qian et al., 2022).

Taken together, these papers define a coherent technical theme for the broader “LightQuant” idiom: correlation structure, induced coherence, and carefully selected observables can replace or augment direct intensity detection under strong background, fading, or jamming.

6. Imaging, nonclassicality, and photonic-state engineering

A second cluster of adjacent work concerns imaging and optical state characterization under photon-starved or otherwise nonstandard conditions. “Low-Light Shadow Imaging using Quantum-Noise Detection with a Camera” reconstructs an opaque object not from intensity contrast but from spatial changes in quadrature variance after a squeezed-vacuum probe interacts with the object. The paper reports reconstruction using a total of 800 photons in the abstract, notes less than one photon per frame on average, and later gives a more conservative estimate of $6\times 10^{-5}$ photons per pixel per frame with 1600 photons in total in the conclusion (Cuozzo et al., 2021). The method is relevant because it shows that imaging information can be encoded in quantum noise rather than brightness.

“White-light Quantitative Phase Imaging Unit” occupies a different position in the same broader space: it is a compact, stand-alone accessory for standard microscopes that combines lateral shearing interferometry with phase-shifting interferometry under white-light illumination (Baek et al., 2016). The reported background phase noise under incoherent illumination is $\sigma = 0.0273$ rad, compared with $\sigma = 0.2462$ rad under coherent laser illumination, and the typical runtime is less than 700 ms for one phase image. “Off-axis holographic imaging with undetected light” extends the undetected-light paradigm to single-shot wide-field amplitude-and-phase reconstruction by combining induced coherence with Fourier off-axis holography in a hybrid nonlinear interferometer; the implementation uses a 405 nm pump, 910 nm signal photons, 730 nm idler photons, an 11.9 ± 0.1 mm field of view, and a resolution of about 5.6 lp/mm (León-Torres et al., 2024).

The supplied literature also includes source and state-engineering work that broadens the operational meaning of quantitative light. “Sub-Poisson-Binomial Light” introduces the experimentally accessible nonclassicality witness $Q_{\rm PB}$ for arbitrary unbalanced multiplexing schemes, with the sufficient nonclassicality criterion $Q_{\rm PB}<0$ and a minimal hardware implementation using a ring resonator and a single on-off detector (Lee et al., 2016). “Silent White Light” models a quantum dot superluminescent diode whose broadband amplified spontaneous emission can be driven from thermal Bose-Einstein statistics toward the Poissonian limit through temperature-dependent quantum-dot occupation dynamics and gain saturation, reproducing the experimentally observed reduction from $g^{(2)}(0)=2$ to approximately $g^{(2)}(190\,\mathrm{K})\approx 1.33$, as reported in the earlier work of Blazek et al. (Hansmann et al., 2023). “Sunlight-Excited Spontaneous Parametric Down-Conversion for Quantum Imaging” then shows that sunlight, filtered around 405 nm and coupled into a PPKTP crystal, can act as the SPDC pump while still yielding position-correlated photon pairs suitable for ghost imaging, with reported coincidence fringes of about 95% contrast (Xing et al., 15 Aug 2025).

At the most abstract end of the spectrum, “Quantum Fluids of Light” provides a many-body framework in which photons acquire effective mass through spatial confinement or diffraction and effective binary interactions through optical nonlinearity (Carusotto, 2022). In that perspective, quantitative use of light is no longer limited to sensing or imaging; it extends to hydrodynamics, driven-dissipative Gross–Pitaevskii dynamics, strongly correlated photonic matter, and topological photonic states.

7. Significance, limitations, and recurring misconceptions

The primary significance of LightQuant in its explicit sense is methodological rather than algorithmic. It is a lightweight, modular framework that organizes Chinese multimodal stock research around curated data, plug-and-play models, and integrated backtesting (Liu et al., 3 Nov 2025). The paper itself implies several limitations: framework specificity to the Chinese stock-analysis setting, dependence on the CSMD-style data model, and a narrower scope than mature general-purpose quant platforms. It is therefore best understood as an infrastructure contribution for research prototyping and benchmark evaluation, not as a universal financial operating system or a novel predictive theory.

In the broader photonic usage, the supplied papers converge on a different set of technical principles. These include correlation-based discrimination against background, induced coherence without direct probe detection, photon-efficient or sub-photon imaging, detector/probe wavelength separation, and explicit characterization of nonclassical statistics. The limitations are likewise concrete and domain-specific: the OPA receiver fails under Rayleigh fading because random phase averaging eliminates the needed phase-sensitive signature (Zhuang et al., 2017); entangled-photon LiDAR prototypes may require long scan times unless acquisition is reduced or scanning is made correlation-driven (Zhao et al., 2021); QuIC LiDAR requires interferometric stability and remains vulnerable in principle to carefully phase-matched jamming (Qian et al., 2022); squeezed-vacuum shadow imaging trades spatial sharpness against bin size and mode-area matching (Cuozzo et al., 2021); white-light QPI units assume sufficiently sparse samples to avoid overlap of displaced images (Baek et al., 2016); and off-axis holographic imaging with undetected light remains limited by low SNR at short exposure times (León-Torres et al., 2024).

A recurrent misconception is that all appearances of “LightQuant” refer to one established interdisciplinary framework. The supplied corpus does not support that reading. Instead, it supports a narrower factual definition centered on the CSMD financial framework and, separately, a broader inferred usage in which “LightQuant” serves as a convenient descriptor for quantitative sensing and imaging with classical or quantum light (Liu et al., 3 Nov 2025). Another misconception is to equate LightQuant with LightQANet; the latter is a distinct low-light image enhancement method with its own codebook-based architecture and training objectives (Wu et al., 16 Oct 2025).

Under this disambiguated reading, LightQuant names either a concrete quantitative-finance framework or, more loosely, a research orientation in which information is extracted from optical fields through structure in correlations, statistics, phase, or induced coherence. This suggests that the term’s long-term meaning will depend less on lexical similarity than on whether future work stabilizes one of these usages into a common reference point.