HARN: Hierarchical Associative Resonance Network for Event-Driven Multi-Timeframe Forecasting
cs.LG · q-fin.CP · q-fin.GN
Quick verdict
Event-driven multi-timeframe forecaster beats PatchTST/TimeXer descriptively; missing persistence baseline and dead code link keep claims unproven.
Summary
The paper introduces HARN, a neural architecture for forecasting financial prices from multiple bar timeframes (e.g., M15/H1/H4) that maintains a persistent state per timeframe and updates each level only when its bar completes, formalizing an event-driven protocol intended to prevent look-ahead leakage. The architecture combines a causal multi-scale convolutional encoder, a gated associative memory with bounded residual-driven writes, all-to-all cross-level 'resonance' attention, and bottom-up evidence aggregation; forecasting is done in basis-point space and reconstructed to prices. On four assets (AAPL, EURUSD, USDCHF, XAUUSD) over three seeds, HARN reports lower reconstructed-price MAE, RMSE, and sMAPE than single-timeframe PatchTST and TimeXer trained on the anchor timeframe (Table 4), but the comparison is admittedly confounded by differing inputs, targets, and training protocols. Ablations (A0-A3) are single-run, cover two assets, and are not parameter-matched; component effects are tiny and A0 fails to match the main results. Crucially, no persistence (zero-BPS) baseline is reported, and the paper's implementation-defined MASE cannot settle whether HARN beats last-value forecasting. HARN's near-zero cross-seed variance (e.g., USDCHF MAE ±0.000000) versus large baseline variance is unexplained. The code audit is static, with no empirical perturbation test; the stated public code URL returns 404 and raw data cannot be shared, so reproducibility rests on unavailable artifacts. The paper is unusually candid, but the empirical evidence remains weak.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
5 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
- HARN maintains persistent per-timeframe representations that update only on completed-bar events, formalizing an event-driven multi-timeframe protocol that avoids look-ahead leakage (Sections 3-4).
- Under the reported configuration, HARN achieves numerically lower reconstructed-price errors than anchor-only PatchTST/TimeXer on all four assets (Table 4), explicitly framed as a descriptive pipeline-level comparison isolating no factor.
- Single-run ablations (A0-A3, Table 5) show no single component dominates, so the independent contribution of the associative memory, resonance, or evidence path is not established.
- A static code audit (Table 6) finds the implementation consistent with the event-level causal contract, subject to the same-event evidence-ordering qualification (row 9).
- The work is positioned as a persistent multi-timeframe framework and protocol contribution, not evidence of universal predictive superiority (Abstract, Section 10).
Strengths
Explicit event-driven protocol formalization
Section 3 (Eqs. 1-5) precisely defines completed-bar alignment, update indicators via timestamp equality, masked convex-combination state transitions, and the causal requirement at event granularity; the batched form (Eq. 4) is shown to coincide with the scalar definition, addressing a real and under-specified problem in multi-timeframe financial ML.
Rare provenance transparency
Table 3 (Section 6.2) maps every reported number to its configuration and explicitly lists what is undetermined; Table 8 inventories divergent defaults across the three trainers; Section 7.1 flags the missing persistence baseline and the omitted percentage reductions rather than hiding them.
Qualified code-level causal audit
Correct statistical framing, even if unexercised
Candor about scope
Weaknesses
No persistence baseline
Confounded headline comparison
Very weak ablation evidence
Reproducibility is nominal, not actual
Unexplained near-zero seed variance
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
HARN (arXiv 2609.26822) is a method paper proposing an event-driven protocol and neural architecture for forecasting financial prices from asynchronous multi-timeframe bar streams (M5-H4). Its central claim is deliberately narrow: HARN formalizes persistent per-timeframe state updated only on completed-bar events, and "the results position HARN as a persistent multi-timeframe forecasting framework rather than evidence of universal predictive superiority" (Abstract). Four claims are made: the protocol formalization (Section 3), the architecture (Section 4), results reported with explicit provenance (Section 6), and a code-level causal audit (Section 9). The scope discipline is unusual and commendable — but it also means the empirical payload is thin by design.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 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
What is HARN?
HARN (Hierarchical Associative Resonance Network) is a neural forecasting model for financial prices that works on multiple bar timeframes at once (e.g., 15-minute, 1-hour, 4-hour bars). It keeps a separate memory state for each timeframe and updates a level only when a new bar completes at that level, so slower timeframes stay frozen between their bar closes — a design intended to avoid look-ahead bias in live trading scenarios.
Does HARN beat PatchTST and TimeXer?
In the paper's runs on four assets (AAPL, EURUSD, USDCHF, XAUUSD) across three random seeds, HARN reported lower price errors than both baselines. However, the baselines were trained only on the fastest timeframe with a different target (normalized close vs basis-point change) and a different training setup, so the author himself calls the comparison descriptive — it does not show that the architecture itself is better.
Is the HARN code publicly available?
Not currently. The paper states that its candidate public URL (github.com/NabeelAhmad9/HARN) returned HTTP 404 when checked during revision, and the raw market data cannot be redistributed for licensing reasons. The author says checkpoints, preprocessing code, and per-seed CSVs exist in a private project repository, but no public release is claimed.
Does HARN actually predict better than a naive 'price stays the same' rule?
The paper cannot answer this. The natural persistence baseline (forecasting zero change, i.e., repeating the last close) was not run, and the reported MASE metric uses a nonstandard denominator. HARN's MASE is below 1 for EURUSD (0.87) and USDCHF (0.80) but above 1 for AAPL (1.33) and XAUUSD (1.14), so even that unreliable metric does not settle whether the model adds value over persistence.
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.
- Figures 1-8 appear only as captions in the extracted text; their visual content (architecture schematics, forecast plots, ablation bar chart) could not be assessed.
- The code-level audit (Table 6) reflects the author's own static review; it cannot be independently verified because no public source is accessible.
- arXiv v1 (Sep 2026); no peer-review status is stated, and the extracted text appears complete (Sections 1-12, Tables 1-10, Appendix A.1-A.5, references present).
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.