Independent Researcher
Financial-systems risk and controls: auditing financial data and automated financial decisions · open datasets and preregistered studies
Previously about 16 years at IDBI Bank, India (2006–2022), Assistant Manager to Assistant General Manager (AGM from 2018).
Research focus. Automated rules that reject credit applications, payments, flagged transactions or trades are hard to evaluate, because outcomes for rejected cases are often missing or incompletely recorded (in credit scoring this is studied as reject inference). Public crypto markets offer a useful setting for studying part of this problem: a rejection can be recorded with the rule that fired, and the asset's later market path observed without executing a trade. Post-Rejection Follow-up Sampling (PRFS) is a documented follow-up procedure for doing this; it draws on about 16 years of banking operations and controls work. Limits: a rejected token's later price is not the return an executed trade would have earned, and it says nothing about how a rejected borrower would have repaid.
All publications, datasets and software, with copy-ready citations →
The monetary penalties on banks that the Reserve Bank of India announced by press release from 2 January 2023 to 8 October 2026: 890 records, of which 885 are bank penalties on 699 banks. Each record is coded for the control area that failed and for how it failed, under a written codebook.
Operational-risk losses of 39 banks as printed in their Basel Pillar 3 OR1 tables, 2016-2025, with the source page and file fingerprint for every value.
Introduces Post-Rejection Follow-up Sampling (PRFS), an observational methodology in which a separate tracking process samples the price and liquidity of each rejected token from the same live data path used by the rejecting system. The companion dataset contains 67,000 follow-up rows across 2,997 rejection events and 457 unique Solana token mints; in these data the deployed tracker sampled each tracked token every 10 minutes and kept one open record per token at a time. The reported estimand is the observed post-rejection forward return; no causal claim about a hypothetical accepted trade is made.
Reports a 15-day paper-traded deployment of an autonomous memecoin trading system on Solana decentralised exchanges (190 trades, 29 March to 12 April 2026). A test of three exploratorily selected weak entry hours against all other hours is directional and not confirmatory (Mann-Whitney p = 0.56), and removing the top three trades (1.6 percent of the sample) makes the cumulative return unprofitable. The contribution is measurement infrastructure, a deposited assertion-based reproduction script, and three design lessons for small-sample trading studies.
arXiv ↗ SSRN ↗ Manuscript concept DOI ↗ Dataset concept DOI ↗ Prospective study: OSF preregistration ↗
Audits eight filter rules in a DEX trading system's filter stack, using about 2,400 rejection events from a two-week window in April 2026. It counts rejected events whose tracked price later fell to half or less of its first tracked value within 24 hours ("saves") or rose to double or more ("misses"); the save-to-miss ratio is descriptive and not by itself evidence of filter skill.
arXiv ↗ SSRN ↗ Manuscript concept DOI ↗ Dataset concept DOI ↗
Public benchmark dataset of algorithmically rejected DEX trading events from a live Solana filter stack: 6,660 rejection events, 169,123 forward-outcome observations and 1,837 lifecycle-tracker snapshots over the window 2026-04-10 to 2026-05-02 UTC. Event-level outcome linkage is available for about half of the events (3,333 of 6,660 exact matches); mint-level coverage is near-complete (1,075 of 1,076 mints). CC-BY-4.0 licensed. Deposited across six platforms (arXiv, SSRN, Zenodo, Kaggle, IEEE DataPort, Hugging Face). RED-2400 is a dataset name, not a count, and is the first window in a planned dataset series.
arXiv ↗ SSRN ↗ Zenodo dataset concept DOI ↗ IEEE DataPort DOI ↗ Kaggle ↗ Hugging Face ↗ GitHub reader ↗
Audit, under a pre-specified analysis protocol, of the graduation labels written by an off-chain data collector for a frozen primary cohort of 749,816 unique pump.fun token mints (12 May to 10 June 2026). The recorded label is the collector's own classification of what it saw while polling a feed of the 50 newest tokens, not the platform's graduation event, and its "timeout" label does not establish non-graduation. A locked logistic association model discriminates well in the 15-day development period (AUROC 0.859) but does not generalise to the following 14 days (AUROC 0.464; 95 percent interval 0.411 to 0.520, which includes 0.5). The dataset's documentation notes that this failure mainly reflects a calendar-spline term in the frozen model rather than evidence that the market changed, and that the last day of the validation cohort is outcome-selected.
Records the later prices of tokens rejected by the eight filters of a paper-trading system on the Solana pump.fun marketplace (June 2026). The deposited corpus contains 496,848 post-rejection price samples across 3,125 rejected tokens.
Asks how far descriptive regularities in data collected from public crypto market interfaces reflect the collection instrument rather than the market. It audits five concurrent feeds spanning 57 days in 2026 (Pyth oracle prices against an OKX benchmark, Solana decentralized-exchange quotes, Aave V3 liquidations and utilization, OKX basis and funding, and Wormhole bridge messages) and derives an audit protocol for data users. The five feeds are deposited as open datasets on Zenodo under CC-BY-4.0: RED-ORACLE-2026, RED-ARB-2026, RED-LIQ-2026, RED-BASIS-2026 and RED-BRIDGE-2026.
SSRN ↗ RED-ORACLE Zenodo ↗ RED-ARB Zenodo ↗ RED-LIQ Zenodo ↗ RED-BASIS Zenodo ↗ RED-BRIDGE Zenodo ↗
Compares the output mint sets of two separately configured observation pipelines from one research programme on the Solana pump.fun launchpad: a cohort-detection pipeline and a rejection-filtering pipeline. In the June-July 2026 window, where the rejection stream covers the whole 17.5-day window, the cohort set contains 623 mints and the rejection set 1,742, with 5 overlapping mints (0.803% of cohort; 0.287% of rejection). In the earlier June window a collection gap leaves the rejection stream covering only 4.4% of the cohort window, so its overlap count is too small to be informative. Only raw counts, observed proportions and Jaccard/Dice indices are reported. The two pipelines' outputs were nearly disjoint, consistent with their different selection rules; the paper proposes reporting cross-collector overlap and collector coverage as a routine check.
Public benchmark corpus of algorithmic filter rejections with post-rejection follow-up samples on the Solana pump.fun ecosystem, 2026-04-11 to 2026-07-16. The 96-day window contains one collection gap (no new rejections from 3 June and no samples from 4 June until 24 June 2026), leaving about 75.6 days of sampling coverage. Corpus size: 716,762 sample records across 4,635 unique SPL token mint addresses and 18,237 distinct rejection events under eight named filter families (shares of sample rows: downtrend_24h 56.13 percent, marketcap_out_of_range 21.38 percent, multi_tf_fade 11.98 percent, timeSlot_not_strong 5.17 percent, hour_blacklist 2.36 percent, ultra_fast_reject_exit 1.64 percent, utc_hour_blocked 0.92 percent, h6_sellers_dominating 0.41 percent).
Python toolkit accompanying RED-2400. Packages the filter-audit workflow as an installable library (source on GitHub; not published on PyPI). The current release (v2) attaches outcomes to each rejection event. License: MIT (toolkit) / CC-BY-4.0 (dataset).
CITATION.cff files that generate BibTeX automatically via the "Cite this repository" button.
An explanatory article on a measurement audit of a public token-launch dataset: what the data measured, and what it did not.
Study plans registered in advance on the Open Science Framework (OSF). Each registration is public, time-stamped and has a DOI.
Replication harness for the RED-2400 benchmark. Release v2 attaches outcomes to each rejection event.
Decision-keyed follow-up schedulers and sampler, plus verification tools (prfs-verify) that measure how much of a rejection registry an outcome log covers. Runs offline with a deterministic simulator; contains no exchange connector, credentials or trading logic. The Software Heritage snapshot below is the v1.0.0 code (April 2026).
SWHID: swh:1:rev:f4571edd99cb1ef6cc41d826cf1b28ffec542bb8 (archived 2026-06-08)
Reader (v0.1.0, May 2026) for the JSON-lines outcomes file; the CSV files of the current RED-2400 deposit load directly with pandas.
Independent public use of the datasets and methods on GitHub:
Three editor-invited manuscript reviews (IEEE Access, September and October 2026; Journal of Trial and Error, September 2026). Seven further public reviews, self-initiated, are listed separately below.
Invited manuscript reviews, completed at the invitation of journal editors:
Self-initiated public reviews (seven): six with DOIs (three on Qeios, three on PREreview) and one on ResearchHub:
Web of Science (ex-Publons, ResearcherID PYY-2705-2026) verifies five of the completed reviews listed above (IEEE Access 1, Journal of Trial and Error 1, Qeios 3) as of 5 October 2026.
Professional memberships:
Standards proposals:
Academic correspondence: arati.kamat@ieee.org. Research collaboration inquiries welcome via my ORCID profile.