Hour-Aware Adaptive Risk Management for Autonomous Memecoin Trading on Solana DEXs: Evidence, Theory, and Design Lessons from a 15-Day Deployment

Arati Uday Kamat

Hour-Aware Adaptive Risk Management for Autonomous Memecoin Trading on Solana DEXs: Evidence, Theory, and Design Lessons from a 15-Day Deployment

T0 review · 3 major / 3 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that a counterfactual rejection tracker can show filter value in memecoin trading, and that the system's profitability is fragile enough that removing three trades erases it.

desk verdict The deposited dataset and audit script are potentially valuable, but the abstract and body report irreconcilable statistics for the same headline test, so the paper's reproducibility claim fails as written; worth referee time only if the authors fix that and add a base-rate control. read the letter →

arxiv 2606.08232 v3 pith:44JZ7KPX submitted 2026-06-06 q-fin.TR q-fin.CPq-fin.RMq-fin.ST

Arati Uday Kamat This is my paper

classification q-fin.TRq-fin.CPq-fin.RMq-fin.ST

keywords autonomoustradingmemecoindecentralisedexchangeSolanatime-of-dayeffectscounterfactualevaluationreturnfragilityrejectiontracking

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish three things about an autonomous memecoin trading system on Solana DEXs, based on a 15-day paper-traded deployment: time-of-day patterns exist but are only exploratory, filter decisions can be evaluated against what happens to rejected tokens, and cumulative profitability is extremely fragile. Its central claim is that the counterfactual rejection-tracking data — 17.9 percent of rejected events lost half their value within 24 hours — shows the filters are net-positive, while the +117.7 percent cumulative return flips negative if the top three trades are removed. A sympathetic reader would care because the method offers a way to measure filter value without risking capital, and because the fragility result is a concrete warning about drawing conclusions from small heavy-tailed samples.

What carries the argument

The central mechanism is the post-rejection forward-sampling subsystem: every token rejected by the hard-gate layer is tracked for 24 hours, and its price trajectory is compared against the earliest available forward sample. This yields event-level 'lost-half' and sample-level 'below-half' statistics that are independent of the accepted-trade cohort. The second piece of machinery is hour-of-day stratification, which assigns each trade to a UTC hour by subtracting the hold duration from the exit timestamp. A small assertion-based audit script (audit.py) is deposited so that every headline statistic can be reproduced from the CSV files, making the measurement infrastructure itself the contribu

What would settle it

Run the deposited audit script against trades.csv and the rejection corpus: a non-zero exit would contradict the reproducibility claim. Separately, if a reanalysis computes the 50 percent drawdown rate among accepted trades and finds it equals or exceeds the 17.9 percent rate for rejected tokens, the 'net-positive filters' conclusion would be unsupported.

Watch

Extended reading notes

Core claim

The paper reports a 190-trade paper-traded deployment in which the streamlined system earned +0.62 percent mean per trade, a 40.5 percent win rate, and +117.7 percent cumulative return, with left skew (-1.21) and excess kurtosis (6.61). It identifies three UTC hours (2, 13, 23) as worst-performing, with a Mann-Whitney U comparison against the rest yielding U=1,274, p=0.22, explicitly exploratory. Its main new measurement is a counterfactual rejection tracker: 4,874 forward samples across 184 rejected tokens show 17.9 percent of events reached a 50 percent drawdown within 24 hours, and 26.0 percent of samples recorded the token below half-reference; the author reads this as evidence that the

Load-bearing premise

The filter-effectiveness conclusion takes the earliest available forward price sample as a stand-in for the price the system would actually have paid, and treats a 24-hour 50 percent drop from that reference as the definition of a 'lost-half' event, without comparing against a control group of accepted or generic tokens.

Editorial extensions

If this is right

If the filter-effectiveness results hold, autonomous DEX traders can evaluate accept/reject decisions by tracking rejected tokens' forward prices, without risking capital.
The exploratory UTC-hour pattern (2, 13, 23) should be treated as a candidate microstructure feature to be re-tested out-of-sample before any operational blacklist is used.
The fragility analysis implies that cumulative return alone is not a trustworthy performance metric for heavy-tailed memecoin trading; distribution-level and concentration statistics are needed.
The precision-recall tradeoff means that adding safety layers can reduce trade flow to zero, so filter strictness must be chosen with the operating frequency constraint in mind.
The deposited audit script allows independent verification of the headline numbers from the raw trade and rejection logs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural next measurement is a control group: computing the same 50-percent-drawdown rate among accepted trades (or a random sample of Solana memecoins) would tell whether 17.9 percent is evidence of good filtering or merely a deadly market.
The reference price used is the earliest forward sample, not a simulated entry price; replacing it with the system's actual execution model would test whether the counterfactual is measuring filter value or just price collapse.
The v1 equity curve was generated from synthetic data and withdrawn; any independent reproduction should rely on the deposited trades.csv and generate the curve by cumulative summation, as v2 does.
If the same methodology were applied to other chains or to longer windows, the time-of-day pattern and fragility concentration could be tested for generalisation.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: the only contribution with real shelf life here is the deposited trade log, rejection corpus, and audit.py, and the paper currently contradicts its own reproducibility promise by reporting two incompatible sets of Mann-Whitney statistics for the same three-hour comparison. If that is resolved, this is a useful data artifact for Solana memecoin microstructure; as written, it cannot be verified.

What is genuinely new: the 190-trade trades.csv and the 4,874-observation rejection follow-up corpus are public measurements I have not seen elsewhere for Solana DEX memecoins. The idea of running a counterfactual tracker on rejected tokens is a legitimate new application of reject-inference, even if the specific method is cited to the author's own companion paper. The paper is also unusually candid: it discloses that the worst hours were selected in-sample, it reports the fragility ablation (top three trades flip cumulative return), and it explicitly withdrew the v1 equity curve generated from a random normal series. That transparency deserves credit.

The soft spots are serious. The abstract gives n=17 for the blacklisted hours, mean -11.60%, U=1,345, p=0.5634; the body and Table 1 give n=18, mean -17.85%, U=1,274, p=0.22. The per-hour counts in Table 1 (5+5+8=18) match the body. No rule produces n=17 from the deposited CSV. Since audit.py is advertised to exit zero iff every headline reproduces, the abstract and body cannot both be reproductions. This is not a minor typo; it hits the paper's principal claim.

The counterfactual filter-effectiveness claim is also weaker than the prose suggests. The reference price is the earliest available forward sample, not the entry price the system would have paid. There is no stop-loss or exit rule applied, and no base rate measured among accepted tokens or the general memecoin universe. Without that control, the 17.9% lost-half rate does not distinguish a good filter from a market where half of any random token loses half its value within a day. The per-category dispersion is reported vaguely; PROVENANCE.md is supposed to have the numbers, but we cannot check because the deposit is not yet accessible.

The hour-of-day pattern is honestly labeled exploratory, so I do not hold that against it. The architecture comparison (streamlined 190 trades vs thirteen-layer 0 trades) is uninformative, and the paper admits that.

Bottom line: this paper is for someone studying memecoin DEX microstructure or building rejection-tracking evaluation tools. The dataset and audit script could be genuinely reusable, but the internal inconsistency and missing baseline mean the conclusions as stated are not yet supported. I would not desk-reject it; I would send it to a referee with instructions to demand (a) a single consistent set of statistics and a corresponding audit rule, and (b) a base-rate comparison for the filter claim. If the authors cannot reconcile the numbers, the paper should be rejected.

Referee Report

3 major / 3 minor

Summary. The manuscript reports a 15-day paper-traded deployment of an autonomous memecoin trading system on Solana DEXs. It deposits a 190-trade log, a counterfactual rejection corpus, and an audit script claimed to reproduce every headline number. The reported results include an exploratory time-of-day comparison (three worst UTC hours vs. the rest), a counterfactual filter-effectiveness analysis, a precision-recall architecture comparison, and a fragility analysis showing that removing the top three trades flips cumulative return negative. The paper is candid about in-sample selection and fragility, but the current v2 text contains an internal statistical inconsistency in the headline time-of-day comparison and draws a net-positive filter-effectiveness conclusion that the deposited data cannot support without a control group.

Significance. The measurement-infrastructure aspiration is genuinely valuable: a deposited trade log, a rejection corpus, and a reproducible audit script would distinguish this paper from the typical non-reproducible DEX-trading report. The author also deserves credit for explicitly withdrawing a synthetic v1 figure and for disclosing the in-sample selection of the three hours. However, the paper is not currently acceptable because the flagship reproducibility claim fails: the abstract and body report irreconcilable Mann-Whitney statistics for the same comparison, so audit.py cannot verify both. In addition, the filter-effectiveness section lacks any baseline or control group, making the 'net-positive' claim unsupported. Both issues are load-bearing and require substantial revision before publication.

major comments (3)

[Abstract vs. §V.A/§VII] The same three-hour comparison is reported inconsistently. Abstract: n=17 for hours {2,13,23}, mean -11.60%, other hours n=173, mean +1.82%, U=1,345, p=0.5634, r=0.085. Body/Conclusion: n=18, mean -17.85%, other hours n=172, mean +2.55%, U=1,274, z=-1.23, p=0.22, r=0.18. Table 1 per-hour counts sum to 18, matching the body, not the abstract. Since only one trades.csv is deposited, audit.py cannot reproduce both sets of numbers. This directly contradicts the paper's claim that the audit script exits zero iff every headline number reproduces. The authors must either withdraw one specification or explicitly define the subset that gives the abstract's n=17. Without this fix, the measurement-infrastructure contribution fails for the paper's central time-of-day finding.
[§V.B (Filter Effectiveness)] The 'net-positive' conclusion is unsupported because no control group or base rate is provided. The 17.9% event-level lost-half rate and 26.0% below-half sample rate are computed against the earliest available forward sample, which is not the price the system would have paid, and no exit or stop-loss rule is applied. The paper never measures the same drawdown rate among accepted tokens or in the general memecoin universe, so the reported rates cannot distinguish filter skill from a generally deadly market. The paper itself concedes in §V.B that some per-category deposit slices do not directly evidence rejections as net-positive. Please add a control group (e.g., accepted-token or matched-market baseline) or abandon the 'net-positive' language and present the rates as descriptive forward-outcome statistics.
[§V.A and §IV.C (Time-of-Day Selection)] The time-of-day comparison is circular in the statistical sense: the three hours (2, 13, 23) were selected as the lowest-mean hours in the same sample used to test them. The reported mean gap (-17.85% vs. +2.55%) is a mathematical consequence of that selection rule, not independent evidence of a market regime. The authors disclose this, which is good practice, but the Conclusion still calls the result a 'candidate microstructure feature' and the abstract presents it as an 'exploratorily identified' pattern. An out-of-sample holdout, a selection-adjusted test, or a pre-registered hour set is necessary before this can be interpreted as a pattern rather than a selection artifact.

minor comments (3)

[Abstract and §VII] The net SOL figure is stated as +0.03873 in the abstract and +0.039 in the conclusion. Please unify the precision.
[§IV.B and §V.E] The paper uses 'cumulative return' as the sum of per-trade percentage returns, not a compounded return. This should be stated explicitly in the abstract and results, as the +117.7% figure could otherwise be misread as a compounded equity return.
[Figure 3 caption] The v1/v2 note is appropriate, but please state explicitly in the version history which earlier analysis numbers, if any, are superseded. The current abstract/body inconsistency suggests the sanitization is incomplete.

Circularity Check

1 steps flagged · score 6.0 of 10

Post-hoc selection of the three lowest-mean hours makes the headline time-of-day mean gap a restatement of the selection rule; filter-effectiveness and internal-reproducibility issues are separate validity defects.

self definitional [§V.A (Time-of-Day Effects), Table 1; echoed in Abstract and Conclusion]
"These three hours were identified as the worst-performing hours by inspection of the per-hour mean P&L breakdown after the data was collected; they were not pre-registered as a blacklist before the trading window began. Comparing the three blacklisted hours (combined n=18, mean -17.85 percent) against all other hours (n=172, mean +2.55 percent) using the Mann-Whitney U test, we obtain U = 1,274, z = -1.23, p = 0.22, rank-biserial r = 0.18."

The group being tested is defined by the outcome it is used to demonstrate: hours 2, 13, and 23 were selected as the lowest-mean hours from the same 190 trades, then grouped as 'blacklisted' and compared with the rest. The reported mean gap (-17.85% vs +2.55%) therefore follows arithmetically from the sorting rule, since the weighted average of the groups with the smallest group means is necessarily below the weighted average of the remaining groups. The MWU p-value is not itself fixed, but the direction and the grouping are, so the test cannot independently support a time-of-day effect; it restates the sample ordering. The caveat that the comparison is exploratory lowers the strength of the claim but does not remove the construction equivalence.

full rationale

The clearest circular step is §V.A: hours 2, 13, and 23 were selected post hoc as the worst-mean hours and then compared with the remaining hours, so the headline mean gap is a mathematical artifact of the selection. This is the main basis for the score of 6, corresponding to a central claim that is partially forced by its own construction. The counterfactual filter analysis (§V.B) is a validity threat rather than a circularity: the 'lost-half' metric uses the earliest forward sample as reference and lacks an accepted-token control group, but 17.9% is not algebraically identical to the 'filters are net-positive' conclusion, so it is better assessed as missing external validation than as a reduction to the input. The abstract/body mismatch in the Mann-Whitney statistics (U=1,345, n=17, p=0.5634 vs U=1,274, n=18, p=0.22) is a serious reproducibility defect for the advertised audit.py guarantee, but it is an internal inconsistency, not a circular derivation. No load-bearing self-citation was found: the pointer to [Kamat 2026b] is a methodological reference, and the deposited data and audit script are the actual evidence base.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or theoretical entities. The load-bearing assumptions are about the validity of paper-trading as a stand-in for real execution, the counterfactual interpretation of rejected-token forward prices, the meaning of p-values after in-sample selection, and the integrity of the deposited data. The free parameters are mostly post-hoc analytical choices (hour selection, reference price, aggregation over undisclosed filter categories).

free parameters (3)

Worst-hour blacklist {2,13,23} = UTC hours 2, 13, 23
Chosen from the same 190-trade sample by ranking per-hour mean P&L (§V.A). The mean contrast to the remaining hours is therefore guaranteed by construction and cannot be evidence for a time-of-day regime.
Reference price for rejection events = earliest available forward sample
The 'lost-half' and 'below-half' metrics are defined against this reference; a different choice (e.g., price at rejection time or would-be entry) would change rates. §V.B.
Six anonymised filter categories / volume-weighting = not disclosed
The aggregate 26.0% is a volume-weighted average over categories whose per-category shares and definitions are not in the paper; §V.B and PROVENANCE.md.

assumptions (4)

domain assumption Paper-trading simulation with 'realistic slippage assumptions' approximates live execution closely enough for the headline return figures.
Section IV.A; VI.A admits real execution would add slippage, failed transactions, and MEV, so the +0.039 SOL net may not transfer.
domain assumption Forward samples of rejected tokens are a valid counterfactual for 'had the system entered'.
The reference price is the earliest forward sample, not an entry price; there is no simulated exit rule. §V.B.
ad hoc to paper Mann-Whitney U p-value remains meaningful after the worst hours are selected in-sample.
The paper reports p=0.22 after selecting hours from the same data; even though it labels the result exploratory, the test's assumptions are violated by the selection, so the p-value cannot be read as a confirmatory quantity (§V.A).
domain assumption Deposited CSVs are complete and audit.py faithfully reproduces the headline numbers.
The abstract and main text report different U/p/n values for the same test, so this assumption is not currently verified (§V.A vs abstract).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hour-Aware Adaptive Risk Management for Autonomous Memecoin Trading on Solana DEXs: Evidence, Theory, and Design Lessons from a 15-Day Deployment." pith.science (2026). https://pith.science/paper/44JZ7KPX

@misc{pith2026260608232,
  author       = {Pith},
  title        = {Pith review of: Hour-Aware Adaptive Risk Management for Autonomous Memecoin Trading on Solana DEXs: Evidence, Theory, and Design Lessons from a 15-Day Deployment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/44JZ7KPX}},
  note         = {Machine review of arXiv:2606.08232}
}

read the original abstract

We report a 15-day paper-traded autonomous memecoin trading deployment on Solana decentralised exchanges (DEXs), designed as a controlled measurement of three microstructure questions on which classical equity theory offers well-defined predictions but on which the AMM Solana venue lacks published measurement: (i) time-of-day return patterns on a 24/7 permissionless venue; (ii) whether decision-time filter stacks are net-positive against counterfactual returns of rejected tokens; (iii) whether small-sample cumulative-return statistics on a heavy-tailed venue are structurally robust or fragile. The 190-trade sample (March 29 to April 12, 2026) shows a 40.5 percent win rate, mean per-trade return +0.62 percent, cumulative +117.7 percent, skewness -1.21, excess kurtosis 6.61. Mann-Whitney U on three exploratorily identified worst entry hours (n=17, mean -11.60 percent) versus all others (n=173, mean +1.82 percent) yields p = 0.5634; directional and non-confirmatory. A parallel counterfactual rejection-tracker collected 4,874 forward-sample observations across 184 rejection events; of 48 events observed for at least six hours, 27 (56.25 percent) reached a 50 percent drawdown from reference (the full-cohort 17.9 percent is a censored lower bound). Removing the top three trades (1.6 percent of sample) flips cumulative return unprofitable. The three findings connect to Kyle (1985) informed-flow, Precup-Sutton-Singh (2000) off-policy evaluation, and Bailey-Lopez de Prado (2014) deflated-Sharpe predictions. Alongside the trade log and rejection-sample corpus (CC-BY-4.0), we deposit audit.py (MIT), an assertion-based reproduction script that exits zero iff every headline number reproduces from the deposited CSVs. Companion dataset: Zenodo concept DOI 10.5281/zenodo.20043301. The paper's principal contribution is measurement infrastructure and three transferable design lessons.

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.