Tabular Deep Learning for Algorithmic Trading: Cross-Regime Bayesian Optimisation for Equity Signal Generation

Joshua Le Grice

cs.LGAug 27, 2026

Tabular Deep Learning for Algorithmic Trading: Cross-Regime Bayesian Optimisation for Equity Signal Generation

Authors: Joshua Le Grice

Organizations: Department of Computer Science University of Exeter

Abstract

Algorithmic trading now represents a market exceeding $20 billion, where even marginal gains in signal robustness can translate into economically significant returns. Existing evaluations of equity prediction models do not explicitly target regime robustness during hyperparameter selection. Five model classes are trained on daily observations from approximately 300 large-cap US equities over eleven years, with Bayesian optimisation configured to target trading performance across three statistically different market regimes. Regime-robust hyperparameter selection is associated with out-of-sample generalisation, as signal precision remains above the random baseline across all four quarters of the test period, and portfolio performance slowly degrades under simulated input noise before collapsing beyond a defined threshold. No individual tabular deep learning architecture outperforms gradient-boosted trees, but combining XGBoost and TabNet using rank aggregation produces a Hybrid ensemble with an annualised return of 51.26%, a Sharpe ratio of 2.44, and a statistically significant CAPM alpha of 0.423 (p = 0.011). A near-zero beta indicates this outperformance is driven by stock selection, not market exposure. Alternative data plays a secondary role once technical and fundamental features are accounted for, as well as contributing more strongly on the short side than the long, and varies by model class. An interactive application makes these results explorable in real time, with live data integration the remaining step toward practical deployment.

Figures & tables

Figure 1: Overview of motivation, methodology, and inference framework

Panel of Figure 2 — Figure 2: Dataset Construction

Panel of Figure 3 — Figure 3: Modelling Methods

Model	Key Architecture	Early Stopping	Hyperparameter Search Space
Logistic Regression	Linear model	–	Regularisation strength $C\in[1\text{e-}5,1\text{e}4]$
XGBoost	Gradient boosted trees	30 rounds	Estimators $\in[1000,1300]$ , learning rate $\in[0.005,0.1]$ , max depth $\in[3,5]$ , min child weight $\in[15,50]$ , $\gamma\in[0.5,2.0]$ , $\lambda\in[0.5,3.0]$ , $\alpha\in[1.0,2.0]$ , subsample $\in[0.7,1.0]$ , column sample $\in[0.5,1.0]$
MLP	Halving layers, LayerNorm, ReLU	10 epochs	Hidden dimension $\in\{128,256,512\}$ , layers $\in[2,3]$ , dropout $\in[0.1,0.3]$ , learning rate $\in[1\text{e-}4,5\text{e-}3]$ , batch size $\in\{2048,4096\}$ , gradient clipping $\in[0.5,2.0]$ , weight decay $\in[1\text{e-}5,1\text{e-}3]$
TabNet	Sequential attention, sparsemax, ghost batch normalisation	10 epochs	$n_{d}{=}n_{a}\in\{8,16,32\}$ , decision steps $\in\{3,4\}$ , $\gamma\in[0.5,3.0]$ , learning rate $\in[5\text{e-}4,1\text{e-}2]$ , batch size $\in\{2048,4096,8192\}$ , gradient clipping $\in[0.5,1.5]$ , weight decay $\in[5\text{e-}4,1\text{e-}1]$
FT-Transformer	Feature tokenisation, multi-head self-attention, linear head	10 epochs	Attention blocks $\in\{2,3\}$ , attention heads $\in\{2,4,8\}$ , embedding dimension $\in\{32,64\}$ , head size $\in\{32,64\}$ , attention dropout $\in[0.2,0.5]$ , feed-forward dropout $\in[0.2,0.5]$ , learning rate $\in[5\text{e-}4,1\text{e-}2]$ , batch size $\in\{2048,4096,8192\}$ , gradient clipping $\in[0.5,1.5]$ , weight decay $\in[5\text{e-}4,1\text{e-}1]$

Table 1: Model Architecture and Training Configuration

		ML Classification Metrics				Backtesting Performance
Year	Model	Acc	Pre	Rec	F1	Ret%	SR	Sor	Cal	DD%
2022	S&P 500	–	–	–	–	-18.99	-0.90	-1.21	-0.78	-24.47
	LogReg	0.792	0.434	0.368	0.369	-6.15	-0.56	-0.52	-0.50	-12.20
	XGBoost	0.799	0.440	0.357	0.350	9.02	0.36	0.76	0.56	-16.01
	MLP	0.794	0.404	0.349	0.337	2.56	0.05	0.38	0.19	-13.76
	TabNet	0.786	0.397	0.353	0.346	7.62	0.33	0.88	0.95	-8.04

Table 2: Model Performance Across All Models and Market Regimes

Risk-Adjusted Outperformance
Statistical Metric	LogReg	XGBoost	MLP	TabNet	FT-Trans.	Hybrid
CAPM $\alpha_{\text{ann}}$	-0.071	0.276	0.208	0.107	0.037	0.423
CAPM $\beta$	0.112	0.200	-0.154	-0.031	0.047	0.048
CAPM $p$ -value ( $\alpha$ vs S&P 500)	0.603	0.171	0.264	0.479	0.833	0.011 ∗
Prob. Sharpe Ratio (PSR)	0.083	0.740	0.499	0.369	0.238	0.960 ∗
Distribution & Significance Tests

Table 3: Statistical Robustness and Distribution Analysis – OOS

Figure 4: Equity Curves of All Models vs S&P 500 Benchmark – OOS

	Gaussian Noise ( $\sigma$ )				Reference
Metric	0.05	0.10	0.20	0.50	Clean	Random Null
JS Divergence (Long)	0.0126	0.0179	0.0243	0.0350	–	–
JS Divergence (Short)	0.0102	0.0129	0.0171	0.0261	–	–
Total Return (%)	25.17	34.36	2.72	6.66	51.26	$-$ 8.21
Sharpe Ratio	1.13	1.53	0.04	0.26	2.44	$-$ 1.44
Max Drawdown (%)	$-$ 7.50	$-$ 9.00	$-$ 14.23	$-$ 9.76	$-$ 7.76	$-$ 8.47

Table 4: Hybrid Model Robustness Analysis: Gaussian Noise Perturbation and Null Hypothesis Test

	Signal Precision (%)				Observations
Quarter	Long	Short	Long vs Baseline	Short vs Baseline	N Long	N Short
Q1 (Jan–Mar)	23.06	26.67	$+$ 13.06	$+$ 16.67	360	360
Q2 (Apr–Jun)	23.12	25.81	$+$ 13.12	$+$ 15.81	372	372
Q3 (Jul–Sep)	25.78	25.00	$+$ 15.78	$+$ 15.00	384	384
Q4 (Oct–Dec)	24.07	24.34	$+$ 14.07	$+$ 14.34	378	378

Table 5: Quarterly Temporal Stability Analysis: Signal Precision Across 2025 Out-of-Sample Sub-Periods

Figure 5: Top 10 features by mean SHAP value for the long and short signals – XGBoost

Model	Signal	Technical	Fundamental	Macroeconomic	Alternative
XGBoost	Long	50.87%	43.73%	3.76%	1.63%
	Short	59.45%	33.30%	4.36%	2.89%
LR	Long	55.93%	14.44%	15.56%	14.08%
	Short	63.09%	11.89%	12.82%	12.20%
MLP	Long	40.94%	32.98%	13.52%	12.56%
	Short	47.51%	27.17%	15.16%	10.16%

Table 6: Total Mean SHAP Contribution by Feature Category Across All Models

Figure 6: Hybrid Ensemble Performance Sensitivity to Position Size – OOS

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: UMAP 2D projection of the feature space coloured by raw next-day return (left) and daily cross-sectional return decile label (right).

Figure 8: UMAP 3D projection of the feature space coloured by raw next-day return (left) and daily cross-sectional return decile label (right).

Regime A	Regime B	KS Statistic	$p$ -value	KS Statistic	$p$ -value
		Return		Volatility (5-day)
2022	2023	0.087	$<0.001$	0.238	$<0.001$
2022	2024	0.101	$<0.001$	0.277	$<0.001$
2022	2025	0.077	$<0.001$	0.178	$<0.001$
2023	2024	0.015	$<0.001$	0.048	$<0.001$
2023	2025	0.020	$<0.001$	0.061	$<0.001$

Appendix

Table 7: Pairwise Kolmogorov-Smirnov Test Results Across Market Regimes.

Figure 9: Kernel Density Estimates of Return and Volatility Distributions Across Market Regimes (2022–2025).

Figure 10: Kernel Density Estimates Comparing Training (2015–2024) Against the OOS Period (2025).

Figure 11: Equity curves for all models vs S&P 500 benchmark, 2022 validation period.

Figure 12: Equity curves for all models vs S&P 500 benchmark, 2023 validation period.

Figure 13: Equity curves for all models vs S&P 500 benchmark, 2024 validation period.

Panel of Figure 14 — Figure 14: Confusion matrices for all models evaluated on the OOS period.

Model	W Statistic	p-value	Normal?
SPY	0.7831	$<$ 0.001	No
Logistic Regression	0.9918	0.177	Yes
XGBoost	0.8739	$<$ 0.001	No
MLP	0.9893	0.062	Yes
TabNet	0.9586	$<$ 0.001	No
FT-Transformer	0.9404	$<$ 0.001	No

Appendix

Table 8: Shapiro–Wilk normality test results for daily returns.

Pair	Ann. $\alpha$	$p(\alpha)$	$\beta$	$R^{2}$
LogReg vs XGBoost	$-$ 0.148	0.226	0.314	0.218
LogReg vs Hybrid	$-$ 0.230	0.059	0.416	0.247
LogReg vs MLP	$-$ 0.082	0.542	0.177	0.059
LogReg vs TabNet	$-$ 0.067	0.621	0.169	0.034
LogReg vs FT-Transformer	$-$ 0.061	0.645	0.236	0.088
XGBoost vs Hybrid	$-$ 0.022	0.894	0.776	0.388

Appendix

Table 9: Pairwise CAPM regression results between models. Ann.

\alpha

is annualised relative alpha.

Figure 15: Hybrid model Gaussian noise robustness analysis. Left: strategy degradation vs noise level. Centre: JS divergence vs noise level. Right: probability distribution shift at $\sigma=0.20$ .

Figure 16: Hybrid model quarterly signal precision across the 2025 OOS period. The dotted line indicates the 10% random baseline.

Model / Direction		Technical	Fundamental	Macroeconomic	Alternative
XGBoost	Long	50.87%	43.73%	3.76%	1.63%
XGBoost	Short	59.45%	33.30%	4.36%	2.89%
Logistic Regression	Long	55.93%	14.44%	15.56%	14.08%
Logistic Regression	Short	63.09%	11.89%	12.82%	12.20%
MLP	Long	40.94%	32.98%	13.52%	12.56%
MLP	Short	47.51%	27.17%	15.16%	10.16%

Appendix

Table 10: Feature Category SHAP Attribution Across All Base Models

Long Signal
Feature	Mean SHAP
Distance from 52-Week High	0.0208
Distance from 50-Day MA	0.0091
Analyst Rating Consensus	0.0065
Relative Volume	0.0064
5-Day Volatility	0.0064

Appendix

Table 11: Top 10 Features by Mean Absolute SHAP Value — Logistic Regression

Long Signal
Feature	Mean SHAP
Market Capitalisation	0.0216
Trading Volume	0.0116
Daily Dollar Volume	0.0086
5-Day Volatility	0.0083
Distance from 52-Week High	0.0068

Appendix

Table 12: Top 10 Features by Mean Absolute SHAP Value — XGBoost

Long Signal
Feature	Mean SHAP
Trading Volume	0.0230
Market Capitalisation	0.0215
Shares Outstanding	0.0132
Distance from 52-Week High	0.0087
Daily Dollar Volume	0.0083

Appendix

Table 13: Top 10 Features by Mean Absolute SHAP Value — MLP

Long Signal
Feature	Mean SHAP
Distance from 52-Week High	0.0131
5-Day Volatility	0.0089
Relative Volume	0.0073
VIX Regime (High)	0.0057
Bollinger Band Width	0.0038

Appendix

Table 14: Top 10 Features by Mean Absolute SHAP Value — TabNet

Long Signal
Feature	Mean SHAP
Market Capitalisation	0.0197
Daily Dollar Volume	0.0161
Distance from 52-Week High	0.0090
Trading Volume	0.0079
Relative Volume	0.0046

Appendix

Table 15: Top 10 Features by Mean Absolute SHAP Value — FT-Transformer

Model	Tuning (30 trials)	Final Training	OOS Inference
LR	32.0 min	22.9 s	0.04 s
XGBoost	10.8 min	8.4 s	0.09 s
MLP	45.0 min	42.3 s	2.18 s
TabNet	19.6 hrs	15.5 min	14.0 s
FT-Transformer	2.9 hrs	3.2 min	8.8 s

Appendix

Table 16: Tuning, Training, and Inference Time by Model

Figure 19: Interactive Backtesting Environment

Panel of Figure 21 — Figure 21: AI Research Assistant Architecture

Algorithmic Trading

Algorithmic trading uses computational models to generate signals, allocate portfolios, and execute orders across financial markets.

See all 149 papers

How do standard time series forecasting benchmarks translate to finance? A broad study of risk adjusted performance
Mar 2, 2026Adir Saly-Kaufmann, Kieran Wood, Jan-Peter Calliess +1Financial Time Series Time Series Forecasting
Machine Learning Research Group, Department of Engineering Science, University of Oxford · Oxford-Man Institute of Quantitative Finance, University of Oxford
Algorithmic Trading Strategy Development and Optimisation
Mar 16, 2026Owen Nyo Wei Yuan, Victor Tan Jia Xuan, Ong Jun Yao Fabian +1Algorithmic Trading
SINGAPORE INSTITUTE OF TECHNOLOGY · INFOCOMM TECHNOLOGY
Deep Reinforcement Learning for Equity Trading: Benchmarking Actor-Critic Methods with Forward Retraining
Sep 25, 2026Bicheng Wang, Xinyi ZhangAlgorithmic Trading Reinforcement Learning
Stanford University

Mar 2, 2026J/K move · Enter open · S save

How do standard time series forecasting benchmarks translate to finance? A broad study of risk adjusted performance

Adir Saly-Kaufmann, Kieran Wood, Jan-Peter Calliess, Stefan Zohren

Machine Learning Research Group, Department of Engineering Science, University of Oxford · Oxford-Man Institute of Quantitative Finance, University of Oxford

Financial Time Series Time Series Forecasting Forecasting Benchmarks Algorithmic Trading Time Series Representation Learning

Open paper AskPDF arXiv

Forecasting financial asset returns is challenging due to low signal-to-noise ratios, non-stationarity and heavy-tailed distributions. We benchmark 16 deep learning architectures for time series forecasting, including Transformers, State Space Models (SSMs), xLSTMs and recurrent hybrids, on a 15-year daily futures dataset using an end-to-end volatility-targeted Sharpe ratio optimisation task. Our results challenge prevailing assumptions in the literature. Several widely used 'state-of-the-art' architectures, including generic Transformers and SSMs, underperformed a simple AR(1) benchmark in a realistic financial forecasting task. The strongest results came from comparatively simple hybrid approaches using recurrent structures: a Variable Selection Network combined with an LSTM achieved the highest Sharpe ratio, while xLSTM showed the greatest robustness to transaction costs. These findings suggest that success on standard time-series benchmarks does not reliably transfer to finance, and that architectures tailored to the unique properties of asset returns may remain more effective than generic attention-based approaches for financial forecasting.

Abstract

Algorithmic Trading

Explore similar work

How do standard time series forecasting benchmarks translate to finance? A broad study of risk adjusted performance

Algorithmic Trading Strategy Development and Optimisation

Deep Reinforcement Learning for Equity Trading: Benchmarking Actor-Critic Methods with Forward Retraining