Quant-Safe Explainable Artificial Intelligence for Dynamic Portfolio Management A Reproducible, Leakage-Resistant ML + SHAP Framework with Execution-Grade Accounting Logs

1. Introduction

The empirical finance literature increasingly applies ML models to return prediction and cross-sectional asset selection [1,2]. However, many reported gains fail to survive realistic deployment conditions due to data leakage, improper temporal splits, and missing trading frictions [3,4]. In parallel, explainable AI (XAI) techniques such as SHAP offer a path to interpretability and regime diagnostics, reducing “black-box” risk in investment decision-making [5,6,7].

This study adapts XAI-driven investment modeling to a stricter engineering standard. Rather than emphasizing headline returns, we emphasize auditability: point-in-time data semantics, leakage prevention, OOS explainability, and execution-consistent performance logging.

1.1. Contributions

We contribute:

• A Quant-Safe architecture defining operational safeguards for financial ML pipelines: reporting lags, point-in-time features, walk-forward evaluation, and OOS-only explainability.
• A formalized Algorithm 1 and a Data Leakage Failure Modes checklist aimed at reviewer scrutiny and reproducible research.
• A naïve vs Quant-Safe comparison table illustrating why common evaluation shortcuts inflate results.
• A reproducible implementation with execution-grade accounting logs (daily mark-to-market equity, realized/unrealized P&L by ticker, slippage proxies) and a Monte Carlo appendix for robustness.

2. Related Work

ML asset pricing and return forecasting have advanced rapidly [1,2]. Yet systematic over-optimism in backtests is well documented, including backtest overfitting [3] and leakage introduced by non-point-in-time features and improper cross-validation [4]. Interpretability frameworks such as SHAP [5] and general XAI guidance [6,7] motivate the use of explanation not merely as a diagnostic, but as a safeguard against spurious relationships.

On the portfolio side, we adopt robust, low-parameter constructions (top-N, inverse-volatility weights) consistent with the preference for simple, stable allocation rules under estimation error [8,9]. We also explicitly model transaction costs, which are central to realistic strategy evaluation [10,11].

3. Data and Experimental Design

3.1. Universe and Horizon

The research universe is the DJI constituent set used in the project codebase. The modeling target is a forward return proxy (as implemented in the pipeline), while the portfolio is rebalanced at a fixed schedule (monthly in the primary configuration).

Limitation (Survivorship Bias): If the study uses a modern constituent list retroactively, survivorship bias may inflate performance. The primary claim of this paper is the Quant-Safe methodology (leakage prevention + OOS explainability), not the absolute historical return magnitude. Future work should incorporate point-in-time constituent membership or investable proxies [4].

3.2. Macro and Risk Controls

We incorporate a minimal macro block designed to be easy to reproduce:

• Interest-rate proxy (e.g., 10Y yield series or yield-change z-scores)
• Risk/volatility proxy (e.g., VIX or realized volatility)

These variables enable regime attribution in SHAP and motivate risk scaling when volatility spikes.

4. Quant-Safe Methodology

4.1. Design Principles

The Quant-Safe architecture enforces:

• Point-in-time semantics: every feature must be available at prediction time.
• Reporting lags: fundamentals are shifted by a disclosure lag to avoid “trading on future filings”.
• Temporal evaluation: walk-forward or strictly OOS evaluation for research claims.
• Explainability discipline: SHAP computed on OOS folds only, preventing explanation leakage.
• Execution consistency: transaction costs, slippage proxies, and accounting-quality logs.

4.2. Algorithm 1: Quant-Safe Pipeline

Algorithm 1 Quant-Safe Pipeline (Walk-Forward + OOS SHAP + Portfolio Layer)

Require: Asset prices $\left\{P_{i,t}\right\}$, macro variables $\left\{M_t\right\}$, optional fundamentals $\left\{F_{i,q}\right\}$ with reporting lag $\Delta$ Require: Prediction horizon H, rebalance schedule $\mathcal{R}$, cost model $\mathcal{C}$, top-N selection rule  1: Align prices to trading calendar; compute returns and technical factors  2: Lag fundamentals: ${\tilde{F}}_{i,t}\leftarrow F_{i,q(t-\Delta )}$  3: Construct feature vector $X_{i,t}\leftarrow [{\mathrm{tech}}_{i,t},{\tilde{F}}_{i,t},M_t]$  4: Label history where forward outcomes are observable: $y_{i,t}\leftarrow \mathrm{Return}(t\to t+H)$  5: for $k=1$ to K walk-forward steps do  6:     Define train window ${\mathcal{T}}_k$ and test window ${\mathcal{S}}_k$ with strict time ordering  7:     Fit model $f_k$ on $\{(X_{i,t},y_{i,t}):t\in {\mathcal{T}}_k\}$  8:     Score OOS: ${\widehat{y}}_{i,t}=f_k\left(X_{i,t}\right)$ for $t\in {\mathcal{S}}_k$  9:     Compute SHAP on OOS only: ${\varphi }_{i,t}\leftarrow \mathrm{SHAP}(f_k,X_{i,t})$ for $t\in {\mathcal{S}}_k$ 10:     Form portfolio at each rebalance $t\in \mathcal{R}\cap {\mathcal{S}}_k$:    Select top-N by ${\widehat{y}}_{i,t}$; weights via inverse volatility; apply caps/turnover constraints 11:     Simulate execution with $\mathcal{C}$; record trades and mark-to-market equity 12: end for 13: Aggregate OOS predictions, performance metrics, and SHAP summaries across steps

4.3. Data Leakage Failure Modes (Reviewer Checklist)

Reviewers frequently reject financial ML work due to unaddressed leakage. Table 1 lists common failure modes and explicit mitigations.

4.4. Naïve ML Backtest vs Quant-Safe Evaluation

Table 2 clarifies why many “state-of-the-art” results do not survive deployment.

5. Modeling and Explainability

5.1. Model Choice

We employ gradient-boosted decision trees due to strong performance in structured financial feature sets and robust handling of non-linear interactions [1,12]. Hyperparameters are tuned within the training window only, consistent with temporal safety [4].

5.2. Out-of-Sample SHAP

SHAP decomposes predictions into additive feature attributions under a cooperative game-theoretic framework [5]. In this study, SHAP values are computed only on OOS test folds to preserve interpretability under real-time constraints and avoid explanatory leakage.

6. Portfolio Construction, Costs, and Accounting

6.1. Portfolio Translation Layer

We map predicted returns ${\widehat{y}}_{i,t}$ into an actionable portfolio:

• Selection: top-N assets by ${\widehat{y}}_{i,t}$.
• Weights: inverse-volatility weights to reduce risk concentration [9].
• Rebalance: monthly (primary configuration).
• Turnover control: trade only when ranks change materially.
• Risk caps: max weight per name; optional regime-based gross reduction.

6.2. Trading Frictions

We include a transaction cost and slippage proxy consistent with execution literature [10,11]. Costs are applied on turnover events; slippage can be approximated via implementation shortfall vs. a reference mid price.

6.3. Accounting-Quality Logs

To ensure auditability, we persist:

• Daily mark-to-market equity curve
• Realized and unrealized P&L by ticker
• Trade blotter with fill reconciliation (broker fills matched back into records)

These records are required for credible research-to-production transfer.

7. Results (DJI Research Validation)

7.1. Performance Summary

The repository includes CSV outputs for equity curves, metrics, and OOS SHAP summaries. Figure 1 and Figure 2 are generated from the included result files.

7.2. Interpretation

A central claim is not “a magic alpha,” but that the explanation traces are economically coherent: rate-related and volatility-related features rise during tightening or stress regimes, consistent with discount-rate intuition and risk repricing.

8. Robustness Appendix: Monte Carlo via Block Bootstrap

Backtests are single-path realizations. We complement point estimates with Monte Carlo resampling of the realized strategy return series using a block bootstrap to preserve short-horizon dependence [13]. We report distributions of CAGR, volatility, and maximum drawdown.

Figure 3. Monte Carlo equity fan chart (block bootstrap). Median path with percentile bands.

Figure 3. Monte Carlo equity fan chart (block bootstrap). Median path with percentile bands.

Figure 4. Monte Carlo distribution summary (CAGR, max drawdown).

Figure 4. Monte Carlo distribution summary (CAGR, max drawdown).

The code to generate these figures and LaTeX-ready tables is provided in the repository (see code/monte_carlo_appendix.py).

Table 3. Monte Carlo block-bootstrap summary statistics for key performance metrics.

metric	p05	p25	p50	p75	p95	mean	std
CAGR	0.054	0.109	0.150	0.192	0.253	0.152	0.061
MaxDD	-0.462	-0.359	-0.303	-0.256	-0.204	-0.314	0.080
AnnVol	0.180	0.191	0.202	0.214	0.234	0.204	0.017
Sharpe	0.356	0.608	0.786	0.980	1.264	0.797	0.276

Table 3. Monte Carlo block-bootstrap summary statistics for key performance metrics.

metric	p05	p25	p50	p75	p95	mean	std
CAGR	0.054	0.109	0.150	0.192	0.253	0.152	0.061
MaxDD	-0.462	-0.359	-0.303	-0.256	-0.204	-0.314	0.080
AnnVol	0.180	0.191	0.202	0.214	0.234	0.204	0.017
Sharpe	0.356	0.608	0.786	0.980	1.264	0.797	0.276

9. Reproducibility: Minimal Code Excerpts

Short excerpts illustrate the reproducibility approach (full code in repository).

9.1. OOS SHAP Discipline (excerpt)

Listing 1: OOS-only SHAP computation (conceptual excerpt)

9.2. Accounting Logs (excerpt)

Listing 2: Execution-grade logs: MTM equity and per-ticker P&L (conceptual excerpt)

10. Limitations and Future Work

Key limitations include survivorship bias (if present), sensitivity to trading frictions, and the need for release-aware macro series. Future work will:

• incorporate point-in-time universe membership,
• expand regime switching (macro stress filters, volatility targeting),
• validate on additional universes (ETFs, ATHEX).

11. Data and Code Availability

Code and results are available on GitHub and archived on Zenodo:

• GitHub: KarmirisP/quant-safe-xai-portfolio
• Zenodo DOI: 10.5281/zenodo.18167108 [14]