Joaquín Milanesio Index How I work Next Contact

Risk has a terrain.

I'm an economist from Córdoba who builds pricing, risk and trading models in R and Python, and I report them the way a referee would read them. I finish my degree in December 2026, and I'm heading for quantitative finance, risk, investment banking and data analysis.

The lines are an implied-volatility surface. Move across it to bend the skew; click to drop a pin and find it again in plate II.1.

Index

Fourteen pieces of work, placed by what they are about and how close they sit to research or to tools. Short of time? Press 90-second route at the bottom of the screen.

Spatial ground

I learned econometrics on land and distance. The first habit it teaches is to test whether neighbouring observations depend on each other instead of assuming they don't. I took that habit into markets.

Should an egg distributor produce its own eggs? In the base case, no.

My thesis appraises vertical integration for a regional distributor, under explicit operating, financing and risk assumptions.

I valued the project by NPV against an assumed real required return of , and used APV to separate operating value from financing. A Monte Carlo of runs draws the inputs from triangular ranges; purchase and sale prices share one draw, which is an assumption, not an estimate.

To place the farm, a conservative GIS screen of ten siting rules cut rural parcels to candidates, which would still need site and permit checks. Weekly demand models were compared on a rolling origin. Completed; defense pending. R, Python, QGIS.

The highest real IRR among the base-case alternatives, against a required return of .calculated

Every base-case alternative has a negative NPV.

What moves with the price of land in Buenos Aires?

A hedonic model of asking prices against area, transport, recorded crime and neighbourhood. Associations, not causes.

For land listings from the city's open data I built the spatial features in R, with QGIS for the maps: thefts and robberies within 500 m, and distance to the subway, the railway and the Obelisco. I added neighbourhood fixed effects, tested the residuals with a global Moran's I, looked for local clusters with LISA, and proposed a SARAR model as the next step.

Adjusted R² of the preferred model.calculated

The base model's residuals are still spatially dependent (Moran's I ), which is why a spatial model comes next.

Spatial Econometrics, May 2026. R (sf, spdep), QGIS.

Fourteen models, one exploratory comparison.

Officially assessed urban land value across the Río Cuarto department, from the province's open parcel data.

Two OLS specifications and twelve other methods (Ridge, Lasso, PCR, PLS, KNN, CART, random forest, GBM, GAM, SVM, XGBoost, CatBoost) on one random split. Some tuning reused the test set, and a random split measures filling gaps in a known map, not predicting new ground.

The lowest test MSE on log value, from XGBoost.calculated

In the random forest, the parcel's coordinates rank first in importance.

Data Science Tools in Economics, December 2025. R.

Which spatial model was hiding in the data?

A course exercise on a simulated process whose true model was hidden. I estimated SLM, SEM, SLX, SDM and SDEM by maximum likelihood. The bottom-up LM tests stopped at a lag model. Testing down from a general nesting model favoured SDEM among the candidates (LR p = , not rejected), which also had the lowest AIC.calculated

Spatial Econometrics, June 2026. R (spdep, spatialreg).

Open the instrument: how a shock spreads

A backtester locked against future prices, used to test a spatial signal.

Stocks in the same sector and supply chain move together. Do the neighbours' returns this week predict mine next week?

The engine moves one event at a time (market data, signal, order, fill), with partial fills, costs and portfolio accounting. Asking for a bar from the future raises an error, and a perturbation test scrambles the future to check the past stays identical. I reconciled it against a vectorised backtest and a toy case done by hand.

The signal is the weighted average of neighbouring stocks' recent returns for 48 large US stocks. The neighbours come from today's sectors and a supply chain I mapped by hand, so this is a retrospective test, not a strategy that could have run at the time. I wrote it down before running it and tested it against a 300-permutation placebo. Python, DuckDB.

Moran's I within the same week, p = , over weeks.calculated

A week ahead there is no evidence of a lead (p = ), and under the assumed costs the strategy loses money.

Markets

Four questions a trading desk would ask. The trading rules are tested on held-back periods or fresh simulated paths, under stated costs, against the simplest thing that could work; the pricer is tested numerically. The simple thing usually holds up.

An option pricer built twice, in R and in Python.

Do Black-Scholes, binomial trees and Monte Carlo agree, are the Greeks right, and what does a real volatility surface look like?

Black-Scholes, CRR trees for European and American exercise, and Monte Carlo with antithetic draws. The analytic Greeks are checked against finite differences. Hull's textbook case reproduces: call , put .

The surface is one day of SPY quotes: Black-Scholes-equivalent vols from American options, with partial no-arbitrage checks. Interpolating along each expiry misses held-out quotes by vol points ( with a whole expiry left out); a single realised volatility misses by . Python, R, DuckDB, Excel.

The largest gap between R and Python across the Black-Scholes prices; Monte Carlo is checked in standard errors.calculated

Click any figure on this page to see where it comes from.

Leaning quotes against inventory halves the swings.

Does inventory control make a market maker earn more, or does it only lower the risk?

I wrote a price-time-priority order book with partial fills, an exact accounting identity and synthetic Poisson flow that includes informed traders. Three quoting rules: a fixed spread, a linear inventory skew and Avellaneda-Stoikov. Parameters are chosen on 300 paths and tested once on 1,000 new ones.

Standard deviation in ticks per 30-minute session with the inventory skew, against for a fixed spread.simulated

It costs a little mean: against ticks, small but real. Avellaneda-Stoikov lands in between (sd ).

Python.

Eight pairs, fifteen years, no evidence of a stable edge.

Does a cointegration-based pairs strategy keep an edge out of sample, after costs?

Eight pairs with an economic reason to move together, 2009 to 2026. Engle-Granger and Johansen in every formation window, a walk-forward over 31 windows, and the last three years held back for the end. Commissions, spread, slippage and borrowing are charged at assumed rates.

USD in the held-back years from trading every pair, trades, after a disclosed bug fix and re-run.backtest

The same rule earned in development. The strict rule traded times.

R.

Seven models, one holdout, no evidence of alpha.

Do machine learning models predict next month's ETF excess returns better than the historical mean?

Nineteen liquid ETFs and thirteen features, including FRED macro data lagged by its usual publication delay. A purged, embargoed walk-forward, and the 2019–2026 holdout run once from a protocol committed beforehand. Placebo and planted-signal controls, with a Holm correction across models.

The best holdout R² against the historical mean, p = .backtest

None beats the mean significantly, even before the Holm correction.

R, SQL.

Credit and macro

Above: the share of household and company credit classified as irregular since 2015, from the central bank's own files. Households in 2025–26 look nothing like the years a model could train on.

Credit models that met a regime they had never seen.

Using only public, entity-level data from Argentina's central bank, can portfolio deterioration be predicted out of time better than naive benchmarks?

I rebuilt the BCRA debtor files for 87 institutions, 2015 to 2026, separately in R and in Python ( rows compared) and checked them against the published PDF. A fractional regression for the irregular share and a logistic model for a deterioration event. Trained to 2021, with 2022 embargoed, and tested once on 2023–2025.

AUC for companies: it ranks which lenders' business portfolios will deteriorate. That is not a default probability.backtest

On the size of the irregular share it loses to persistence: RMSE against percentage points. By July 2026, of household credit was classified irregular.

R, Python.

Money and prices in Argentina, 1993–2024.

Monthly CPI, monetary base, activity and interest rates from BCRA and INDEC. ADF and Phillips-Perron tests, Engle-Granger and an exploratory error-correction model. It was my first econometrics paper: the equation does not test the quantity identity it starts from, so I left the hypotheses inconclusive. Today I would rebuild it in price levels and growth rates.

Econometrics I, 2024. EViews.

Open the instrument: why unit roots come first

Tools

Software I use myself: a workbench for portfolios and a logbook that keeps score of my own calls.

A portfolio workbench where every strategy has to beat 1/N.

Mean-variance and minimum-CVaR frontiers, Black-Litterman and hierarchical risk parity, all run through the same walk-forward after costs and judged against equal weights with a bootstrap and a deflated Sharpe ratio. It also has Fama-French and local factor models, stress tests on Argentine crises, DCF valuation, a forecast ledger scored out of sample, and daily feeds from BCRA, INDEC, FRED and SEC EDGAR.

Python, SQLite, JavaScript.

A logbook that keeps score of my own predictions.

Every call on a company is sealed with a date and a condition that would prove it wrong. When the date arrives I check it, and the errors I confirm become calibration lessons. It sits on a universe of 21,000+ firms with competitor and supply-chain links, plus an ETL of SEC EDGAR statements.

Python, JavaScript.

Last-Mile Atlas

An internal platform I built at IG Networks that maps last-mile providers and territory economics across Latin America. It is company work, so it stays a one-line mention.

How I work

Four rules I hold the six quant projects to; the coursework above came before some of them. The fourth is the one people skip, so it's the one you can try yourself below.

  1. Write the plan before seeing a result

    The question, the metric, the benchmark and what would change my mind go into a dated file first. Changes to the plan are dated and kept.

  2. Beat the naive answer or report that it didn't

    The historical mean, persistence, equal weights, a single volatility. If the model loses to them, that's the result.

  3. Compute it twice

    Key numbers are rebuilt independently: R against Python, code against the published source.

  4. Keep the future locked

    Walk-forward with purge and embargo where it applies, held-back periods run once, engines that throw an error when they peek. A re-run after the test is disclosed, never hidden.

About authorship: I design each project, meaning the plan, the choice of models, the assumptions and the equations. The code, the runs and the documentation were done with AI assistants (Claude, Codex) to go faster, and each repository says so.

The unsurveyed part of the map

Where I'm going next, and the questions my own projects left open.

  1. Now

    From pricing networks to pricing risk

    Today I quote and price last-mile connectivity across Latin America at IG Networks. Next I want to take the same habits into finance: quantitative analysis, risk, investment banking, or data and spatial analysis, wherever a number has to hold up once someone checks it.

  2. Dec 2026

    Thesis defense and degree in Economics

    National University of Río Cuarto. The thesis is written; the defense is next.

  3. 2027

    Master's in Quantum Finance, University of Buenos Aires

    Quantum amplitude estimation for option pricing and tail risk, and quantum optimisation for portfolios. The options engine on this page is the classical benchmark I'll measure them against.

  4. Open

    Questions my projects left open

    Bring the pairs and ML strategies into the event-driven engine so all of them run under one set of rules. Rebuild the supply-chain network from dated, sourced filings. Test whether credit deterioration spreads between lenders the way prices spread between neighbours.