Review Papers Against Their Own Results

 Quick Explanation Copied

 Long Explanation

Core Claim and Supporting Evidence

Yang and Yang propose three hybrid forecasters—LSTM-TFE, LR-TFE, and BR-TFE—that decompose a series with Ensemble Empirical Mode Decomposition (EEMD) into IMF components plus a residual, train one model per component, and sum the predictions EEMD hybrid forecasting paper. This follows the well-established noise-assisted decomposition logic of Wu & Huang's original EEMD Original EEMD method paper, and the reported gains are consistent with later EEMD-based hybrids such as EEMD-attention BiLSTM load/price forecasters EEMD+attention electricity forecasting comparison.

The strongest reported improvements are on the two real series. For temperature, LR-TFE cuts RMSE from 1.393 to 0.713 (−49%); for SP500, from 14.17 to 7.30 (−49%), with R2 rising to ~0.9995. LSTM-TFE reduces SP500 RMSE from 559.6 to 478.8 (500 iterations). BGPT computes these percent reductions directly from the paper's Tables 3 and 4:

Note the log scale: LSTM on SP500 (RMSE 559.6) dwarfs the LR-family bars; the hybrid consistently halves error within each model family.

Critical Weaknesses and Blind Spots

• Data leakage risk in decomposition. EEMD is applied to the whole series before the 6:4 train/test split; IMF boundaries computed over test-period data constitute look-ahead information. This is a known pitfall in decomposition-forecasting hybrids and is not addressed in the paper—a rigorous rolling-window decomposition (as done for STL in other work Leakage-safe rolling decomposition example) would likely shrink the claimed gains.
• No cross-validation, significance testing, or seeds. Single splits of three series (SP500 n=1,500 test points), no error bars, no repeated runs—EEMD is stochastic (noise ensembles), so reproducibility depends on unreported parameters.
• Synthetic result is nearly vacuous. The sine test is trivially predictable (LR baseline R2 = 1.0), so it cannot discriminate methods.
• Implausibly high R2 on SP500 (0.9995) for next-day index prediction is a red flag: one-step-ahead persistence-like predictions on near-random-walk indices can inflate R2 without economic value; no transaction-cost or directional-accuracy evaluation is provided (contrast with horizon-aware financial benchmarks Financial index ML benchmarking).
• Error fusion is naive summation; per-IMF errors can compound, and no ablation isolates the EEMD contribution versus mere windowing effects (later EEMD work includes such ablations Ablation contrast in later EEMD work).

Verdict

The empirical trend—halved RMSE from EEMD decomposition—is real as reported but rests on a leakage-vulnerable design, three series, and no uncertainty quantification. Treat the improvements as suggestive, not established. What would falsify the claim: re-running with rolling decomposition and strict chronological validation on diverse series; if gains collapse to noise-level, the method's value is overstated.

 BGPT Paper Review

Study Novelty

EEMD + per-component ML prediction was already well established by 2020 (e.g., EEMD-ELM wind-speed and wind-power hybrids cited in the paper itself); the contribution is a straightforward combination, not a new mechanism.

Scientific Quality

Clear structure and reproduced tables, but no cross-validation, no significance tests, no EEMD ablation, whole-series decomposition before splitting (leakage risk), implausibly high SP500 R2, and a trivial synthetic benchmark that cannot discriminate methods.

Study Generality

The divide-and-conquer decomposition idea generalizes across domains, but only three series (one synthetic, two real) are tested and generalization to volatile or longer-horizon series is admitted by the authors to be poor.

Study Usefulness

Provides a replicable recipe (EEMD + LR/BR/LSTM per IMF + summation) with downloadable public data (UCI, Yahoo Finance), useful as a starting baseline, though practitioner value is limited by missing deployment/testing rigor.

Study Reproducibility

Data are public and equations are given, but EEMD noise parameters, LSTM architecture sizes, learning rates, and random seeds are unreported, and EEMD's stochasticity makes exact replication impossible without those details.

Explanatory Depth

Mechanistic explanation is textbook-level (EEMD sifting, LSTM gates) rather than insight-generating; why EEMD improves forecasts is asserted ('milder subsequences') without analysis of error propagation through the sum-fusion or comparison to alternative decompositions.