A Hybrid LSTM-XGBoost Framework for Multi-Horizon Stock Return Prediction Across Diversified Equity Portfolios
Seif ElDein Mostafa, Yahia Ahmed, Farah Datwish, Marwa Solayman
cs.AI
Quick verdict
Proposes hybrid LSTM-XGBoost for multi-horizon stock returns; results show no directional skill vs naive baseline and only 0.0002 RMSE gain over XGBoost-Only.
Summary
The paper proposes a two-stage hybrid architecture for predicting cumulative returns of 14 US equities across six sectors over 30/90/252/365-trading-day horizons. A 2-layer stacked LSTM (64 hidden units) compresses 60-day windows of five scaled market features into 64-dimensional embeddings, concatenated with 14 lagged technical indicators into a 78-dimensional vector feeding per-horizon XGBoost regressors. Training pools all stocks with strict chronological 70/15/15 splits and per-stock MinMax scaling fitted on training data. On the 30-day horizon the hybrid attains global test RMSE 0.0949, about one-third of the standalone LSTM's 0.2799, but only 0.0002 below the XGBoost-Only baseline (0.0951), with wins split roughly evenly across stocks; all R² values are negative, so no model beats a mean predictor. The paper's most distinctive move is a naive always-positive baseline audit: directional accuracy rises to 97.6% at 365 days but sits below the base rate at every horizon (e.g., 57.0% vs 59.9% at 30 days), so no directional skill is demonstrated. A composite investment score produces June 2016 BUY/SELL ratings from unvalidated arbitrary weights. Key limitations: no significance testing (only recommended), internal inconsistencies between Tables III and V for the same configuration, a single bull-market sample ending mid-2016, and no code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 3
● low 0
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The hybrid LSTM-embedding + XGBoost model achieves the best global 30-day test RMSE (0.0949), roughly one-third the standalone LSTM baseline (0.2799).
- LSTM temporal embeddings add selective, sector-dependent gains over XGBoost-Only, winning per-stock RMSE on about half the stocks, with only a 0.0002 global improvement.
- Long-horizon directional accuracy up to 97.6% at 365 days is a base-rate artifact; the model never beats a naive always-positive predictor at any horizon.
- All models show negative R², meaning none beats a mean predictor of return magnitude.
- A composite multi-horizon investment score can rank the 14 equities into BUY/SELL tiers, e.g., STRONG SELL for NVDA and AMZN as of June 2016.
Strengths
Naive-baseline audit of directional accuracy
Table VII and Section IV-D show the hybrid sits below an always-positive predictor at every horizon (gaps of −2.9 to −8.0 pp) and explicitly instruct readers to treat 97.6% at 365 days as base rate, not skill — unusual candor for this literature.
Leakage-conscious protocol
Strict chronological 70/15/15 splits per stock (Section III-A), per-stock MinMax scalers fitted on training data only, and all 14 technical indicators delayed by one day (III-B) to prevent look-ahead.
Transparent negative results
Full specification supporting reimplementation
Weaknesses
Marginal gain over the only meaningful baseline
Internal inconsistencies in reported numbers
No demonstrated skill on either axis
Statistical testing deferred, not performed
Stale, unvalidated investment layer
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper trains a two-stage pipeline on daily OHLCV data from a public Kaggle dataset [20] for 14 US equities across six sectors, January 2010 onward, with the test period ending June 2016. A 2-layer stacked LSTM (64 units, 60-day windows, five scaled features) is frozen after training; its 64-dimensional final hidden state is concatenated with 14 one-day-lagged technical indicators into a 78-dimensional vector, feeding per-horizon XGBoost regressors that predict cumulative returns over 30/90/252/365 trading days. Training pools all stocks with chronological 70/15/15 splits and per-stock MinMax scaling fitted on training data only. Central claims: (1) best global 30-day RMSE of 0.0949, about one-third of the LSTM-Only baseline; (2) selective, sector-dependent gains from temporal embeddings; (3) a composite investment score for portfolio ranking; and — unusually honestly — (4) long-horizon directional accuracy (97.6% at 365 days) is a base-rate artifact, with the model never beating a naive always-positive predictor.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does this model actually beat the market or make profitable predictions?
No. The paper's own numbers show all R² values are negative (no model beats a simple mean predictor of returns), and directional accuracy sits below a naive always-positive baseline at every horizon (for example 57.0% versus 59.9% at 30 days, 97.6% versus 97.8% at 365 days). The authors state a transaction-cost-adjusted backtest would be required before any profitability claim, and none is performed.
What is the hybrid LSTM-XGBoost architecture described in the paper?
A two-layer stacked LSTM with 64 hidden units reads 60-trading-day windows of five scaled market features and outputs a 64-dimensional embedding (its final hidden state). This embedding is concatenated with 14 lagged technical indicators into a 78-dimensional vector that is fed to an XGBoost regressor, trained separately for each prediction horizon of 30, 90, 252 and 365 trading days, on pooled data from 14 US stocks.
What data does the paper use and how recent is it?
Daily open/high/low/close/volume data for 14 US equities across six industry sectors, taken from a public Kaggle dataset, from January 2010 onward. The test period ends around June 2016, so all results predate major regime changes such as Nvidia's GPU-driven growth phase, which the authors themselves flag as a limitation.
Is code available for this paper?
No code repository is provided. However, the paper specifies the full architecture, hyperparameters, search grids and training schedule (Tables I and II), and the underlying data source is a public Kaggle dataset, so the pipeline could be reimplemented from the text.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure-embedded diagram labels (Figs. 1–2) were extracted as interleaved text fragments; all captions, tables and equations are intact and usable.
- The supplied metadata lists an author as Farah Datwish while the paper text reads Farah Darwish; the paper text was treated as authoritative.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.