Bolgaç
Gülen
I build trading systems, run them on live capital, and audit my own results until they break or hold.
Open to quant research, quant developer and trading-desk roles anywhere in Denmark. The fastest way to test me is a conversation: bolgacg1@gmail.com.
Nordea Markets interviewed me for the Desk Quants team in January 2022, while I was finishing my MSc; in the follow-up call, as I remember it, the decision came down to me and one other candidate, and the difference was programming. The subject was never the problem: my 2020 bachelor project was an algorithmic trading system. So I spent the years after in industry and my own robotics company, built the engineering half, and now I run it every day.
Seven results killed in a day
By early July my research systems were reporting extraordinary paper returns across an always-on fleet. On 9 July 2026 I spent one day finding out which of those numbers were real. Almost none were.
A desk's most expensive failures are the quiet ones: strategies that report profit because of a bug nobody went hunting for. So I hunted, and this section is the evidence that I catch those in my own work first.
The venue is a prediction market, chosen for one property: ground truth. Every position resolves to a known outcome, and every fill, meaning every executed trade, is public, so a claimed result can be checked against what actually happened. The tools are the same ones a desk runs on any asset class: order books, the public queue of buy and sell orders, executed fills, calibration, cost models, risk limits.
I spent the day recomputing every headline number from the raw trade ledgers. Almost nothing survived. Every debunked number traced to a specific, nameable bug, and this is the table I wrote that day. The recomputation is now a standing gate, not a one-off: a result does not count here until it survives it.
| Result | What it claimed | The bug |
|---|---|---|
| mirror engine | Millions realized | A bookkeeping rewrite flipped each position's side and price but not the cash that moved, so the ledger booked profit that never existed. Killed |
| bot-06 | Triple-digit return | Simulated stop orders filled at the price the model hoped for, with no check of what the market could supply, and open exposure several times the account it started from. Killed |
| model-trainer-v2 | Best trained model | The best of 214,195 tested combinations, a winner by search luck, trading at prices with no costs. Killed |
| Neural nets | Multiples of capital | Look-ahead: the models saw future information, and putting entries in time order ruined every one. Killed |
| copium | Mispriced low-probability outcomes | Hand-picked subsets of a few dozen trades; across the full universe the same rule lost money. Killed |
| over-dog | Best of 7,800 models | The winners in training and the winners in testing were different models; only 4 of 7,800 held up in both. Killed |
| “riskless arbitrage” | Riskless spread | Two stale prices on a market with almost no volume in it at all. Killed |
Source: estate audit, 9 July 2026. Every claim above was recomputed from the raw trade ledgers before it was struck out.
Then the honest scoreboard
The engines built honestly from the start are the other half of the story. Measured with realistic costs, where every trade had to buy at prices actually available in sizes actually available, they lost. All of them. That is the finding rather than the failure, and the sharpest case was an engine that won the great majority of its individual trades and still gave back most of its capital: once costs are real, the share of trades that win stops predicting whether a strategy makes money.
I keep those results in front of me rather than behind me. They are the reason the recomputation is now a standing gate instead of a one-off: a result does not count in this program until it has survived being recomputed from the raw ledgers by someone trying to kill it, and that someone is me.
Every position resolves to a public outcome and every executed trade is recorded, so a claim can be audited rather than trusted. That property is why the program ran there: it is a laboratory where being wrong is provable, which is exactly what a research method needs and what most venues cannot offer.
Whether any of this survives at size, at real speed, without becoming the adverse selection, the easy counterparty everyone else trades against, is not yet answered, and I say so. Anyone can show a green curve. The skill a desk actually needs is the one that spent a day turning its own green curves red, on purpose, and could tell you exactly why each was wrong.
Live capital, hard limits
Real money went live on 10 July 2026, behind one pre-registered rule on a deliberately small account. Pre-registered means the whole rule, entry, size and exit, was written down and frozen before the first trade, so I cannot bend it after seeing results.
The rule is deliberately narrow: a single entry condition in a short window before an event resolves, on markets liquid enough to fill it, held to resolution, one position per market and no discretionary exit. It trades a pocket where fees are near zero, so if an edge exists there it is measurable rather than eaten by costs. It is also the single forward trial the sealed study in the method section allowed: that study's overall verdict was negative, one narrow pocket earned exactly one test, and the stake is sized for the possibility that the answer is no.
The capital base is deliberately small. The demonstration is the discipline, not the size: every safeguard below was wired in before the first dollar, and each is built to overrule me.
Fixed position size
Every position is a fixed fraction of current equity, compounding, set before the first trade.
Sanity tripwire
Orders that fail basic checks are blocked before they ever reach the market.
Permanent auto-stop
If the account falls past a drawdown line fixed in advance, the system halts for good and cannot restart itself. Profit and loss is read from a ledger of confirmed trades, never guessed from order status.
The live figure is roughly +11 percent to date, and I will not headline it: the sample is small, and the correct posture on a young live number is patience.
Starting small is how bugs get found before they get expensive. Trial runs on an earlier bot surfaced a sizing bug that quadrupled position sizes and a fee feed reporting the wrong rate; both fixes are built into every engine that came after.
What it demonstrates is the discipline: a pre-registered rule, live capital, and a limit that stops me before I can argue my way past it.
Staying honest
The hard part is not finding a signal. Search hard enough and the data will always hand you one. The hard part is telling the real ones from the artifacts of your own searching.
Every method below exists because it once caught me.
What a check is worth can be measured: the two cards below are the same strategies scored with and without an honest clock, and the gap is the whole argument.
01Pre-registration
Before a test I write the rule down and freeze it, so I cannot move the goalposts after seeing results. The calibration study, a test of whether market prices misjudge probabilities by more than fees, ended in one sealed shot: a small pre-registered sample in which the great majority of individual trades won and the fee-adjusted mean was still negative, with a confidence interval spanning zero. A strategy that wins most of its trades and still loses after fees is not an edge, so I reported it as inconclusive-negative and put no real money behind it.
02Cluster bootstrap by market
Positions in the same market tend to win or lose together, so they are not independent evidence. Resampling by market instead of by trade stops a cluster of correlated positions from posing as a large sample.
03One family, one false-discovery rate
Test enough candidate rules and some will look brilliant by pure luck. Benjamini-Hochberg control treats the whole sweep as one family, which separates a real result from the thousands of cells the sweep threw up by luck alone.
04Walk-forward with an embargo
Train on the past, test on the future, in strict time order, with a gap between them so an outcome near the boundary cannot leak backward. Look-ahead is the most expensive bug in this field and the cheapest to introduce; the first card above is what it costs, measured directly.
05Causal replay
Rerun every historical trade as it would have fired live, with only the information available at that instant. The second card above is this check doing its job: the apparent edge in the esports study did not survive an honest clock.
06Two-source discipline
A result confirmed on one data source is a hypothesis, not a finding. A stop-loss variant scored clearly positive on one measurement setup and clearly negative on an independent source, so I rejected it. The precise instrument and the wide-coverage archive answer different questions.
07Peak decomposition
A signal measured at the bottom of a price dip quietly borrows from the rebound that defines the bottom. Separating the two caught a fake edge in cricket markets and flipped a ranking of esports titles I had been ready to believe.
The through-line
I try to break my own results before the market does it for me, and I write down the break so the next run inherits the check.
The research program
One program produced the audit, the method and the live account above. It started with scanners, tools that watched live prices and flagged promising patterns. They found plenty.
I retired the scanners after two days: nothing they flagged could be checked against a real outcome, and an unverifiable signal is not a result.
Luck disguised as skill came next. A study I codenamed copium searched hundreds of thousands of rule combinations across tens of thousands of markets with known outcomes, testing the favourite-longshot bias: crowds systematically overpay for longshots. Every profitable pocket it found was a fluke of that much searching, and a second project running thousands of models confirmed the pattern: winners picked on past data did not repeat.
The fix was one honest engine: it charges realistic trading costs and rejects trades the market could not actually have filled. In paper trading under those rules, every strategy lost money. That was the finding, not the failure: a system can win most of its trades and still lose money, and the starkest case is in the audit above.
In May the program stopped trying to forecast outcomes directly and turned to what could be observed instead, which is the direction the surviving work still runs in.
What the program delivers is not a book of winners. It is a method that builds something exciting, proves it wrong before real money touches it, and turns each kill into a check that runs automatically the next time.
The machinery
The recorders come first, because data you did not capture yourself is data you cannot trust. The results on this site are checked against market data these machines captured themselves, and they are recording right now.
The whole operation runs on two repurposed laptops that stay on 24 hours a day, built and operated alone. One carries the always-on fleet of recorders and trading engines; the other runs a Danish property-auction monitor unattended. No cloud, no cluster: everything is written to run reliably on small hardware.
The hardest lesson was operational. Under load, the main laptop kept dropping its network connection, and the obvious reading was a network fault. The operating system's own load counters said otherwise: the machine was starved for processing power and memory, and the fix was to shed workload, not touch the network.
I change nothing on that machine without confirming the diagnosis first. That is the same rule the research runs on: measure before acting.
Recording at that scale needed its own engineering. A custom on-disk format cut storage by more than half, and separate recorders corroborate the main one: a recorder on Kalshi, a US-regulated exchange, for cross-venue comparison, and a poller reading the public record of completed trades. A full experiment replay finishes in under a second on a two-core box.
The engines that trade on this data are built to be pessimistic. Their simulators charge realistic costs and reject any trade the market could not actually have filled. A simulator that flatters you is worse than none.
All of it is version-controlled: 1,168 commits in 2026 across 15 repositories, counted rather than claimed, each a time-stamped record of a change. That history is the audit trail.
Data systems
The same discipline produces finished systems outside trading. Each project here is a complete data system: a carefully chosen source, a model built on it, and a service that keeps itself running unattended.
Watches the Danish forced-auction market. A tvangsauktion is a court-ordered property sale, and every property has an official assessed value on record. The service tracks each listing, scores its price against that value, flags second auctions, where no minimum bid applies and prices can fall furthest, and alerts over Telegram. It reads tvangsauktioner.dk, the official gazette's aggregator, chosen because it is the legally safe source, and runs unattended on a home server.
Keeps my backtests honest about volume. Volume is recorded after the fact, and borrowing it from the future is the classic way a backtest flatters itself. volmodel models how volume builds toward an event, fitted on 1,894 clean pairs and checked against an independent source it never saw: on the 15 to 60 minute window it was validated for, estimates stay within a factor of about 1.24 of the truth, off by roughly a quarter. The kind of internal data tool a desk actually runs: unglamorous, load-bearing, validated before anything was allowed to depend on it.
Searches a photo library three ways: by colour with an eyedropper, by text description, or by visual similarity to a reference photo. The last two run on CLIP, a pretrained model that maps images and text into one space so they can be compared directly. It runs entirely on-device with no cloud calls, and it is the clearest single proof that I can take a pretrained model and ship a working retrieval system around it.
Shipped products
Finished software, launched and maintained, not demos. I take an idea to a running artifact and keep it running after launch; these projects, none of them about markets, are the proof from outside trading.
minimalist-workout-app
A workout tracker I designed, built and shipped alone, launched on the App Store in July 2026: fully offline, no account required, with crash reporting and a two-track rollout: updates go to a test device before they go wide. Software people run on their phones has to be operated, not just launched; the trading fleet runs on the same discipline.
mocap
Turns phone video into hand and arm motion data you can inspect in an interactive 3D viewer.
video2scene
Builds a 3D model of a scene from a phone video, on an ordinary computer with no graphics card.
islamic-geo & girih
Two engines that draw Islamic geometric patterns from classical construction rules, each an interactive browser gallery.
Different domains, one habit: understand the method, write the tool, make it run.
About & contact
I am an engineer by training and a quantitative researcher by the work I chose to do next. I took an MSc in Engineering, Innovation and Business at SDU in Sønderborg, finishing at 21, which I was told made me the university's youngest Master's graduate, after a fast-tracked BSc in Mechatronics Engineering.
The quant thread starts there: my 2020 bachelor project was an algorithmic stocks trading system, Python news scraping and sentiment analysis run through QuantConnect, with a recession-prediction model on FRED data beside it. During the degrees I also co-founded Alviss Capital Research, a small student research group with no capital of its own, where I ran commodities and cryptocurrency analysis, published a report, and taught the newer members Python. Those two threads are the bridge between the Nordea interview and everything on this site.
Since then I founded Black Flag Robotics, building fixed-wing drones with a focus on flight systems, onboard imaging and cost-competitive manufacturing, through the EU-funded Beyond Beta accelerator. Before that I worked in sales coordination and key account management at Dinex and in product development engineering at Cabin Denmark, and I have held an investment property in Odense since 2022.
The engineering discipline in this portfolio comes from building physical systems under real constraints; the quantitative rigour comes from wanting to know whether a number is true.
Open to quant research, quant developer and trading-desk roles, anywhere in Denmark. I am based in Odense, work in English and Danish (Prøve i Dansk 3), and can meet on short notice.
Contact
- bolgacg1@gmail.com
- Phone
- +45 93 98 12 13
- Location
- Odense C, Denmark
- linkedin.com/in/bolgaç-gülen