Tallinn — MMXXVIOpen to rolesFinancial research · Risk · Analytical systems
Daniil
Daletski.
Testing financial ideas—and building the systems that make them usable.Five years ago I joined a 24/7 service floor; today I plan its capacity, answer for its availability KPI and sign its monthly numbers. Along the way came a finance degree with honours and the research and reporting systems this site documents — each tested the way models should be: out of sample, with the limits stated.
- Tallinn, Estonia
- 5+ yrs · 24/7 operation
- RU · EN · ET · DE
Career · Studioworks OÜ
Promoted through three roles in five years.
2026
What this step added
- Capacity plan ownership
- Availability KPI
- Monthly reporting model
- Hiring decisions
Capability · evidence flow
Every domain traced to the work that proves it.
THThesis2021–2025 empirical study
NXNEXUSNative research terminal
ETEvaluation toolDaily operational performance
OPOperationsStudioworks, 2021–present
ACCourseworkCFA-aligned risk, VaR exercise
Languages
- 01Русский
- 02English
- 03Eesti
- 04Deutsch
Working stack
Analysis
- Python
- SQL
- BigQuery
- Advanced Excel
Modelling
- XGBoost
- Random Forest
- Elastic Net
- Panel OLS
- Tree SHAP
Risk
- VaR
- Expected Shortfall
- Monte Carlo
- Stress testing
Build
- TypeScript
- Rust
- React
- Tauri
01 / Achievement index · verified outputs
Six proof points.Each opens to its evidence.
Academic distinction, empirical scale, a model result, a contractual availability KPI, native software and a deployed product—each one connected to the underlying evidence further down this page.
Evidence available01DegreeAcademic distinctionCum laude
EBS · International Business Administration
↗02ThesisEmpirical scale25,876firm-week observations
↗03Figure 6Verified model result0.2146same-week OOS R²
↗04OperationsOperational availability99.84 %across 49,800 planned hours
↗05NEXUSNative system96views · 67 Tauri commands
↗06ProductDeployed deliveryLIVEdaily operational performance evaluation tool
↗02 / Systems
Three systems. One discipline.
Independent applications of the same method: collect evidence, test or control the decision, and leave an auditable output. Each one states where it stops as loudly as what it achieved.
State
ResearchPublished study, reproducible
Paper stateInstalled build, no live execution
LiveDeployed and in use
What it proves
End-to-end ownership from raw data to a validated, interpretable model
Native software that carries model evidence into a controlled decision surface
A shipped product with a human-review gate designed into the workflow
Headline result
0.2146same-week OOS R²
96registered views
10screenshots per request
Stack
PythonSQLBigQuery
TypeScriptRustTauriReact
Full-stackScheduled deliveryControlled access
Can online attention improve weekly return models?
Yahoo Finance, Finnhub, Google Trends, NewsAPI and FinBERT / Loughran–McDonald scoring. ≈130,000 outbound calls.
0.2146same-week OOS R²
25,876firm-week observations
99U.S. + European firms
Boundary
At a one-week-ahead horizon, no model produced robust out-of-sample predictability. Reported explicitly; not presented as a trading rule.
03 / Professional record
Five years running the operationthat the numbers describe.
The research and the software sit on top of a working operations job: a 24/7 service floor where the capacity plan, the availability KPI and the monthly report are mine to own, defend and reconcile.
Studioworks OÜ · Tallinn, Estonia
Shift Manager — Operations, Workforce Planning & Analytics
2021 — present- 01Front-line operationsService delivery inside the 24/7 cycle.
- 02Senior operational roleQuality, escalation and shift coordination.
- 03Shift ManagerOperations, workforce planning and analytics.
01234567890123456789+years in a 24/7 operation2021 → present
0123456789012345678901234567890123456789employees concurrently scheduledpeak concurrent coverage
~0123456789012345678901234567890123456789FTE workforce base plannedHR-sourced
0123456789012345678901234567890123456789service lines · two portfoliosplanning scope
0123456789012345678901234567890123456789.0123456789012345678901234567890123456789%Combined availabilityContractual KPI met
- Agreed service hours
- 49,900
- Delivered
- 49,819
- Unplanned downtime
- 81 h
Jan–May 2026 planning cycles. Downtime is attributed by driver rather than absorbed into a single figure.
Coverage & supply — Demand becomes service hours, service hours become people.
- Lead a 24/7 operation with up to 61 employees concurrently scheduled across two portfolios and up to 19 service lines, balancing demand against staffing supply, skill certification and rotation constraints.
- Translate demand into service-hour and FTE requirements — up to 10,900 service hours per month, 49,800 across the Jan–May 2026 cycles — reconciled line by line against agreed service time and planned closures.
- Convert the HR-sourced FTE base into coverage and quality ratios per service line, quantifying effective staffed hours against nominal headcount to expose gaps before they reach service.
- Own the handover to Scheduling: raise coverage and feasibility constraints and keep every plan change traceable to an approved assumption.
Availability & downtime — The contractual KPI, and the 81 hours it absorbed.
- Deliver 99.84% combined availability across 49,900 agreed service hours, meeting the contractual availability KPI and holding unplanned downtime to 81 hours.
- Own the escalation and downtime register: 170 incidents across 11 monthly cycles, 168 with timed start-to-fix durations, each classified by cause.
- Rebuilt the incident resolution-time KPI as formula-driven output from source records, clearing 48 pre-existing formula errors.
- Separated downtime frequency from downtime impact: the most frequent fault category and the costliest were not the same — and redirected mitigation to the smaller, costlier driver.
Productivity assumptions — Assumption drift is a capacity risk before it is a variance.
- Test the productivity and AHT assumptions behind the plan against realised data every cycle — required service hours and FTE scale inversely with realised throughput.
- Identified a systematic drift across five consecutive cycles in a core planning input, raised it as an assumption problem rather than a one-off variance, and used quality and downtime data to separate causes.
Data & KPI governance — Every published figure traceable back to a source row.
- Own 10-sheet monthly reporting across two portfolios and 10+ KPIs, consolidating five sources in a 3,400+ formula-cell model with eight defined KPI steps.
- Validate 140,000+ records over five cycles, excluding duplicates and voided entries and resolving cross-sheet inconsistencies before publication.
- Govern a 55-code error taxonomy covering 450+ coded exception records and 190 voided transactions split by manual and system cause.
Evaluation & hiring — Scoring that stays comparable between managers.
- Run structured performance evaluation against an anchored 0–5 scale with a defined observation minimum and 10 assessments per employee per month, reconciling scoring across evaluators.
- Interview candidates and contribute directly to final hiring decisions, assessing skill and operational judgement against current and forecast coverage needs.
- Assess the staffing impact of demand and supply changes — line openings and closures, absence spikes, training lead times, attrition.
- Report assumptions, risks and trade-offs to Scheduling, HR, IT and operational leadership.
Demand becomes service hours, service hours become people.
61concurrently scheduledagainst a ~97 FTE planned base
19 service lines · two portfolios
Split shown by portfolio grouping, not by line volume.
9,960cycle average10,900monthly peak
49,800 service hours planned across the Jan–May 2026 cycles.- Lead a 24/7 operation with up to 61 employees concurrently scheduled across two portfolios and up to 19 service lines, balancing demand against staffing supply, skill certification and rotation constraints.
- Translate demand into service-hour and FTE requirements — up to 10,900 service hours per month, 49,800 across the Jan–May 2026 cycles — reconciled line by line against agreed service time and planned closures.
- Convert the HR-sourced FTE base into coverage and quality ratios per service line, quantifying effective staffed hours against nominal headcount to expose gaps before they reach service.
- Own the handover to Scheduling: raise coverage and feasibility constraints and keep every plan change traceable to an approved assumption.
Plan vs actual · KPIs owned and reported
What the data showed, and what changed because of it.
Service availability
99.84%across 49,800 planned hours
Contractual KPI met
→Downtime attributed by driver, not absorbed into one figure
Throughput & AHT assumptions
5consecutive cycles
Systematic drift — not one-off variance
→Re-based the service-hour requirement; flagged capacity risk
Effective staffed hours
Gapnominal vs deliverable
Headcount overstated available hours
→Coverage adjusted per service line before it reached service
Incident resolution time
168of 170 timed at source
The data existed; the KPI did not use it
→KPI rebuilt formula-driven; auditable to source
Built in-house · in production use
Three systems the operation now runs on.
Monthly Operational Performance & Availability Model
Advanced Excel · two portfolios · up to 19 service lines
10sheets~2,900calc rows8KPI steps2portfolios
Eight numbered steps from total possible hours through availability, FTE and coverage per line, to error rates and mean resolution time.
Reconciled monthly against five source feeds and reported plan-vs-actual with variance commentary.
Company-wide Employee Performance Platform
Python · SQL · web application · 136 employees
5data domains4user groups136employeesRBACaccess
Self-initiated. Evaluations, errors, attendance, attitude and skills in one place, with role-based access and database-backed logging.
Replaced scattered manual records with a single auditable source for absence, error and skill data.
Quality Evaluation Framework & Training Tooling
HTML / JS · Excel
6weighted categories21incident types0–5anchored scale
The framework the operation is scored against, plus an offline scenario trainer and a monthly schedule-review matrix.
Standardised how quality and coverage data enter the reporting cycle.
Additional experience
Southwestern Advantage — Texas, USA
Sales Representative · Summer 2024
Ran an independent sales territory through a 12-week field programme, planning daily activity from conversion data and adapting targets weekly in a fully autonomous international environment.
04 / Research dossier
Online Information Flowand Short-Term Stock Returns.
Linear and machine-learning evidence from the U.S. and European markets. Estonian Business School bachelor’s thesis, 2026 — an auditable pipeline that turns raw price, attention and news data into interpretable models, then validates them under a strict chronological design.
Research objective
Test whether abnormal search activity, news intensity and financial-news sentiment add incremental information to short-horizon equity-return models after controlling for market, volatility, liquidity, size, value, momentum and sector effects.
0123456789012345678901234567890123456789large-cap firms50 U.S. · 49 European
0123456789012345678901234567890123456789,012345678901234567890123456789012345678901234567890123456789firm-week observations2021W01 → 2025W52
0123456789012345678901234567890123456789model-ready variablessignals + controls
01234567890123456789independent data sourcesmarket · search · news
Chronological governance
- 2021 — 2023Train
- 2024Tune, then freeze
- 2025Evaluate once
The panel · figure
Every firm-week in the study, in chronological order.
Train 2021 — 2023Validate 2024Test 202599 firms × 261 weeks · 25,876 retained observations
50 U.S.49 European
The 2025 block is touched once, after the model is frozen. Nothing to its left is re-fitted afterwards—which is what makes the out-of-sample figure below an out-of-sample figure.
Selected out-of-sample evidence · same-week horizon
Where the machine-learning gain is real — and where it is not.
Pooled OLS benchmark (L1)XGBoost (ML1)Out-of-sample R² · axis to 0.25
Full panelh = 0
0.1624
0.2146
+5.2 ppDiebold–Mariano 7.13, p < 0.001 — the gain survives the test
United Statesh = 0
0.1740
0.2237
+5.0 ppStrongest regional result across the study
Europeh = 0
0.1479
0.1489
Incremental machine-learning gain not statistically significant
ASVI was the highest-ranked information-flow feature in SHAP attribution.
Thesis · Figure ML-A · SHAP attribution
The same feature, four regimes, three different signs.
Pushes the prediction upPushes it downSHAP contribution to NVDA weekly return · bps
+28
W0820 Feb 2025Rally, rising attention
+5
W2230 May 2025Sell-off, weak ASVI
-18
W3101 Aug 2025Panic-search regime
+35
W4507 Nov 2025Rally, rising attention
A pooled linear model has one ASVI coefficient and must pick a single sign. Attention behaved differently in a rally than in a panic-search week—and reproducing that is where the machine-learning gain in the table below actually comes from.
The boundary this research reports
At a one-week-ahead horizon, no model produced robust out-of-sample predictability.
Forecasting models did not outperform the mean-return benchmark. The thesis states this explicitly and does not present the result as a standalone trading rule. Same-week explanatory power and next-week forecasting power are different claims, and only one of them survived the test.
End-to-end capability
- 01Acquire≈130,000outbound callsAcross five external sources.
- 02Engineer13variablesASVI, news intensity, FinBERT and Loughran–McDonald sentiment.
- 03Model4specificationsPooled OLS benchmark against XGBoost, Random Forest and Elastic Net.
- 04Validate≈2,916grid fitsOne evaluation on the frozen 2025 window.
- 05AttributeSHAPper observationTree SHAP decomposition with stage-level audit logs.
Model-risk discipline
No look-ahead
Every predictor is available by the close of week t. Nothing in the feature set could only be known later.
Frozen pre-processing
Scaling and encoding are fitted on training data only, then frozen for validation and test.
One shot at the test window
A 243-setting hyperparameter grid is selected on 2024 validation alone; the 2025 window is touched once.
Negative results retained
Insignificant and regime-sensitive outcomes stay in the thesis instead of being trimmed out of the story.
- 19-module Python pipeline, plus TypeScript, Rust and SQL infrastructure
- Fixed seeds, cached inputs, version-pinned release and stage-level audit logs
- Full reproducibility set reruns from a clean clone in ≈70 minutes
Data sources
Yahoo FinanceFinnhubGoogle TrendsNewsAPIFinBERTLoughran–McDonald05 / Capability
The full toolkit,and what each part has been used on.
Four working domains. Every one of them is attached to something on this page that was actually built, tested or reported — not a self-assessed proficiency bar.
Every row is a claim. Every mark is where that claim was actually exercised. Select a body of work to see only what it proves.
24 capabilities · 5 bodies of work
01
Financial risk
6Value at Risk & Expected ShortfallHistorical and parametric, $100k position
Not claimed against ThesisNot claimed against NEXUSNot claimed against Evaluation toolNot claimed against OperationsProved by CourseworkStress & scenario testingRegime splits and shocked inputs
Not claimed against ThesisProved by NEXUSNot claimed against Evaluation toolNot claimed against OperationsProved by CourseworkMonte Carlo simulationSimulated loss distributions
Not claimed against ThesisNot claimed against NEXUSNot claimed against Evaluation toolNot claimed against OperationsProved by CourseworkModel assumptions & limitationsDocumented failure modes, stated boundaries
Proved by ThesisProved by NEXUSNot claimed against Evaluation toolNot claimed against OperationsProved by CourseworkInterest-rate, FX, equity & commodity riskExposure types and derivatives
Not claimed against ThesisNot claimed against NEXUSNot claimed against Evaluation toolNot claimed against OperationsProved by CourseworkRegulatory readingCRR/EBA, IRRBB/CSRBB, Basel IV/FRTB — self-directed
Not claimed against ThesisNot claimed against NEXUSNot claimed against Evaluation toolNot claimed against OperationsProved by Coursework02
Quantitative methods
6Panel OLS, double-clustered SEFirm- and week-level clustering
Proved by ThesisNot claimed against NEXUSNot claimed against Evaluation toolNot claimed against OperationsNot claimed against CourseworkXGBoost · Random Forest · Elastic Net243-setting grid, ≈2,916 fits
Proved by ThesisProved by NEXUSNot claimed against Evaluation toolNot claimed against OperationsNot claimed against CourseworkChronological backtestingTrain → tune → freeze → evaluate once
Proved by ThesisProved by NEXUSNot claimed against Evaluation toolNot claimed against OperationsNot claimed against CourseworkDiebold–Mariano & bootstrappingForecast-comparison significance
Proved by ThesisNot claimed against NEXUSNot claimed against Evaluation toolNot claimed against OperationsNot claimed against CourseworkTree SHAP attributionPer-observation feature decomposition
Proved by ThesisProved by NEXUSNot claimed against Evaluation toolNot claimed against OperationsNot claimed against CourseworkScenario & sensitivity analysisWhat-if on inputs and assumptions
Proved by ThesisNot claimed against NEXUSNot claimed against Evaluation toolProved by OperationsNot claimed against Coursework03
Engineering & data
6Python19-module pipeline; performance platform
Proved by ThesisProved by NEXUSProved by Evaluation toolProved by OperationsNot claimed against CourseworkSQL & BigQueryExtraction, transformation, aggregation
Proved by ThesisNot claimed against NEXUSProved by Evaluation toolProved by OperationsNot claimed against CourseworkTypeScript & RustTauri shell, 67 native commands
Not claimed against ThesisProved by NEXUSNot claimed against Evaluation toolNot claimed against OperationsNot claimed against CourseworkReproducible pipelinesFixed seeds, version pinning, audit logs
Proved by ThesisProved by NEXUSProved by Evaluation toolNot claimed against OperationsNot claimed against CourseworkAPI collection, validation, reconciliation≈130,000 calls; 140,000+ records validated
Proved by ThesisNot claimed against NEXUSProved by Evaluation toolProved by OperationsNot claimed against CourseworkAdvanced Excel modelling3,400+ formula cells, eight KPI steps
Not claimed against ThesisNot claimed against NEXUSNot claimed against Evaluation toolProved by OperationsProved by Coursework04
Operations & workforce analytics
6Capacity planning & demand-to-FTEUp to 10,900 service hours per month
Not claimed against ThesisNot claimed against NEXUSNot claimed against Evaluation toolProved by OperationsNot claimed against CourseworkAvailability & shrinkage modelling99.84% across 49,900 agreed hours
Not claimed against ThesisNot claimed against NEXUSNot claimed against Evaluation toolProved by OperationsNot claimed against CourseworkAHT & throughput assumption testingMulti-cycle drift identified at source
Not claimed against ThesisNot claimed against NEXUSNot claimed against Evaluation toolProved by OperationsNot claimed against CourseworkPlan-vs-actual & variance root cause10 sheets, 10+ KPIs, five source feeds
Not claimed against ThesisNot claimed against NEXUSNot claimed against Evaluation toolProved by OperationsNot claimed against CourseworkSLA & KPI governanceAuditable to source, reproducible monthly
Not claimed against ThesisNot claimed against NEXUSProved by Evaluation toolProved by OperationsNot claimed against CourseworkTeam leadership & hiringUp to 61 concurrent; interviews and selection
Not claimed against ThesisNot claimed against NEXUSNot claimed against Evaluation toolProved by OperationsNot claimed against CourseworkDelivery & leadership
The part of the job that is not a model.
06 / Primary evidence
Open the underlying work.
Real figures and real application screens take priority over portfolio decoration. Every plate opens at full resolution.
Open artefact ↗Thesis · Figure 7 · p.48
Explainability, case by case
Four NVDA firm-weeks show attention changing sign and magnitude across market contexts.
Open artefact ↗NEXUS · Installed build
Risk Command
Exposure, Greeks, VaR, CVaR and historical stress views in a research prototype.
Open artefact ↗Thesis appendix · Figure A1
System architecture
The boundary between external evidence, cloud processing, local compute and audit concerns.07 / Academic distinction
Cum laude
International Business Administration
Estonian Business School · 2023–2026
Finance-focused degree culminating in an empirical thesis that connects market attention, machine learning and an implemented analytical system.
Risk coursework · CFA-aligned
Verify on LinkedIn ↗- Value at Risk
- Expected Shortfall
- Stress & scenario analysis
- Monte Carlo methods
- Model assumptions & limitations
Additional risk foundation · Disney, $100k position
Value at Risk is a quantile, not a number.
248 daily returns · Apr 2024 — Apr 2025
95%$2,56899%$6,409
−8%−4%0+4%+8%
Reads the loss straight off the observed distribution — the 5th and 1st percentile of what actually happened.
The two methods disagree, and the disagreement is the finding.At 99% the historical figure is $6,409 against a parametric $4,424. The observed left tail is fatter than the normal assumption allows, so the parametric model understates the loss by roughly $2,000 exactly where it matters most.
Academic exercise on 248 daily returns from $100,000 notional; not a current investment view.Recognition
Winner — Baltic Essay Contest in GermanAwarded for written argument in German — his fourth working language — alongside a 12-week autonomous sales programme run entirely in English in Texas.08 / Console
A working miniature of the terminal.
The panels carry the same verified figures published across this site; the VaR and valuation calculators run right here in the page on the formulas shown. No market feed, and nothing represents a live position.
NEXUSportfolio editionpaper statesrc 5 feeds · 25,876 obsno live data
2021train2024tune2025test
99firms
261weeks
25,876obs
Full panel
0.2146
United States
0.2237
Europe
0.1489
Linear benchmark vs XGBoost · axis to 0.25
hist $16,022param $11,060248 daily returns · Apr 2024 → Apr 2025
−8%−4%0+4%+8%
$16,022historical VaR
$11,060parametric VaR
$12,646expected shortfall
Daily · scales the published Disney study to your position · not advice
$92intrinsic value / share
years 1–5 · $23terminal · $69
r \ g∞2.00%2.50%3.00%8.0%$101$109$1179.0%$86$92$9810.0%$75$79$84
Two-stage DCF · 5y explicit + Gordon terminal · your inputs only — not advice
+28W08
+5W22
-18W31
+35W45
ASVI flips sign across four 2025 regimes
NEXUS · portfolio edition — published figures, plus calculators that run right here.
Type `help`, or select a module on the left.
$↑Tab↵
Panel OLS·Double-clustered SE·XGBoost·Random Forest·Elastic Net·Diebold–Mariano·Tree SHAP·Out-of-sample R²·Value at Risk·Expected Shortfall·Monte Carlo·Stress testing·Python·SQL·BigQuery·TypeScript·Rust·React·Tauri·Advanced Excel·Demand-to-FTE·Capacity planning·Plan-vs-actual·KPI governance·
Panel OLS·Double-clustered SE·XGBoost·Random Forest·Elastic Net·Diebold–Mariano·Tree SHAP·Out-of-sample R²·Value at Risk·Expected Shortfall·Monte Carlo·Stress testing·Python·SQL·BigQuery·TypeScript·Rust·React·Tauri·Advanced Excel·Demand-to-FTE·Capacity planning·Plan-vs-actual·KPI governance·
Panel OLS, Double-clustered SE, XGBoost, Random Forest, Elastic Net, Diebold–Mariano, Tree SHAP, Out-of-sample R², Value at Risk, Expected Shortfall, Monte Carlo, Stress testing, Python, SQL, BigQuery, TypeScript, Rust, React, Tauri, Advanced Excel, Demand-to-FTE, Capacity planning, Plan-vs-actual, KPI governance.
08 / Financial risk · Quantitative research · Workforce & capacity analytics · Decision systems