Олександр

OLEKSANDR DOVHULIA
Systems & Data Analyst | Data Integrity, Validation & Experiment Design

Kyiv, Ukraine (Remote) • [відкрити контакти](див. вище в блоці «контактна інформація») • github.com/OleksanderSS

SUMMARY
Systems & Data Analyst specialising in data integrity, point-in-time validation, and research reliability. Five years of
experience pricing pre-match and live sports markets at Favbet, followed by hands-on development and auditing
of an AI-assisted market-data research pipeline. Identifies look-ahead leakage, survivorship bias, silent record
loss, and historical data distortions; validates fixes using automated tests and known-answer controls.

EXPERIENCE

Independent Research — Market-Data Pipeline & Point-in-Time Research Core

Sep 2025 – Present | Remote

Technologies: Python, SQL (DuckDB), pytest, Git
AI-assisted engineering: Claude Code, Gemini, Codex; architecture, technical specifications, automated validation
and verification. 1,390+ repository commits.

Phase 1 — Pipeline development and audit

● Built and audited a seven-stage pipeline covering market prices, SEC filings, FRED macroeconomic data,
news, options, and analyst estimates. Under realistic trading costs, no tested strategy outperformed the
no-trade baseline; this negative result was retained rather than optimised away.
● Look-ahead leakage: Found date-only values interpreted as midnight timestamps across three data
sources. In the audited SEC dataset, 46% of 10-K/10-Q facts appeared available before EDGAR acceptance.
Corrected availability timestamps using EDGAR acceptance times: acceptance timestamps were recovered
for 5,804 of 5,805 (accession, filed) pairs (99.98%). A separate known-answer control confirmed that
the acceptance date matched the filed date for 99.87% of filings accepted before 17:15 ET.
● Survivorship bias: A long portfolio of 105 manually selected stocks, short the market as a benchmark,
generated 6.86% annual excess return without a predictive model (Sharpe ratio: 1.22; t-statistic: 6.3;
1996–2023). The result demonstrated how universe selection can create apparent performance without a
predictive model.
● Silent data loss and price distortions: Identified a cleaning step that removed 14,598 valid price

observations, including observations from market-crisis and earnings days. Found that three-class targets
with labels {-1, 0, 1} were incorrectly inferred as regression targets. As a result, 308 models (154 models
across two targets) were trained and evaluated using R² instead of classification metrics; median R² was
−0.0042. Also identified split-adjusted prices that distorted per-share commission estimates by
approximately fivefold (9.7 vs. 2.0 bp); corrected the underlying calculation and the previously reported
figure.

Phase 2 — Evidence-driven redesign

● Replaced the fixed stock list with a test fixture rather than a candidate universe for strategy discovery.
● Built a reusable point-in-time research core that timestamps data by public availability and includes
delisted companies. Added known-answer controls for measurements, hypothesis preregistration with
attempt tracking, and a hash-chained forward journal that evaluates forecasts only against subsequent
outcomes. The system has tracked approximately 500 companies daily since 29 September 2026.
● Automated operations: Built 14 nightly collectors covering news, options, analyst estimates, SEC filings
and 8-K events, job postings, FRED data, and small-cap stocks. The collection cycle runs in approximately
2.5 hours. Added 939 contract tests, a CI gate against new test failures, historical drift monitoring, a
morning report, and scheduled AI analyst runs governed by documented constraints.
● Statistical power before data collection: Estimated statistical power before registering a hypothesis. For a
US-market hypothesis on ageing-related sector effects, estimated power ranged from 0.42 to 0.53 under an
optimistic information-coefficient assumption of 0.10; at an IC of 0.05, estimated power was 0.21–0.32. The
study was not registered or run because the plausible post-decay effect was below IC 0.10, leaving
expected power below the predefined 0.50 threshold.
● Survivorship bias in index backtests: Backtesting current S&P 500 constituents against historical index
composition produced apparent annual outperformance of 5.11 percentage points for April
2018–September 2026 and 5.60 percentage points for January 2013–October 2017. The estimated
contribution from using future index membership was 3.12 and 5.44 percentage points, respectively. The
comparison used equal weighting and monthly rebalancing.
● Historical data revisions: For one industry signal, the agreement between first-release and revised FRED
data was 0.53, showing that revised historical values may materially change the evidence available to a
researcher.

Sports Trader — Favbet

2020–2025

● Priced and managed pre-match and in-play betting markets against defined turnover and margin targets;
benchmarked tennis prices against global market makers in real time.
● Served as the primary market maker for partner-studio esports events without external market prices to
benchmark against. Compiled in-house match statistics and developed pricing models from scratch; the
resulting data feed was licensed to external B2B bookmakers.
● Promoted based on margin performance while maintaining zero critical pricing remarks. Identified mirrored

odds and pricing latency before they could be exploited.

Quality Department — Fozzy Group

2020

● Performed incoming quality control, storage inspections, and documentation audits under corporate
standards and HACCP requirements.
● Conducted sampling and failure analysis using FMEA and prepared non-conformance reports.

PORTFOLIO
Market-data audit — github.com/OleksanderSS/market-data-audit

Documented findings supported by reproducible tests and audit evidence; reviews of two open-source
trading-research projects (FinRL and Stock-Prediction-Models); a failed forecast and a previously reported result
corrected after an audit.

Lead processing MVP — github.com/OleksanderSS/lead_pipeline_mvp

Built a FastAPI webhook that normalises landing-page form submissions, generates a Gemini LLM summary,
classifies leads as HOT, WARM, or COLD using explicit rules, stores results in Google Sheets, and notifies the
team via Telegram. Includes 23 tests and CI on every push.

SKILLS
Data & Research: Data validation, look-ahead leakage audits, point-in-time data, survivorship bias, first-release vs.
revised data, corporate actions (splits and dividends), experiment design, statistical power, preregistration,
known-answer controls.

Technical: Python, pandas, SQL (DuckDB), pytest, Git, REST APIs.

AI Engineering: AI coding-agent orchestration, automated validation, and guardrails (Claude Code, Gemini,
Codex).

Quality: FMEA.

EDUCATION & LANGUAGES
Master’s degree in [official English degree and field from diploma supplement] — National University of Food

Technologies, Kyiv | 2015–2020

Languages: Ukrainian (native); English (B2, professional working proficiency)