Five Years of Losing Money So You Don't Have To

I've been trading since 2021. My first year of university. I had my first allowance, some saved-up pocket money, and the kind of confidence you only get when you don't know anything.

That was five years ago. I'm graduated now (December 2025), and I'm still at it. Still losing sometimes. Still learning. The winning was sporadic. The losing was consistent. That's the worst kind of relationship with trading — the occasional win keeps you hooked just long enough to lose more.

I wasn't a trader. I was a gambler with extra steps. And I didn't even know it.

What University Taught Me (That Didn't Work)

I studied actuarial science. Statistics, probability theory, stochastic processes, time series analysis, differential equations. All the fancy concepts that sound great in textbooks. And I tried to apply every single one of them to trading. Of course I did. That's who I am.

ARIMA models — Looked amazing on historical data. Completely useless live. Markets aren't Gaussian. They're not even close. The stationarity assumption is a fantasy. Markets evolve, adapt, and break your models.

GARCH for volatility — Captures volatility clustering, sure. But capturing volatility isn't predicting direction. I'd know the market was about to move, just not which way. Useful information? Yes. Tradeable edge? No.

Copula theory — I asked my lecturer to go deeper on this. Everyone else wanted to move on, so I learned it myself. The theory is decent. The implementation is a nightmare. Correlation between assets changes in a crisis. Copulas assume a static relationship. Markets don't do static. They do chaos.

Jacobian multipliers and continuity conditions — Beautiful math. Almost entirely useless for actually making money. It's like knowing how an engine works but not knowing how to drive.

The brutal truth: most academic statistics doesn't apply to trading. Not because the concepts are wrong, but because the assumptions don't hold. The distributions aren't normal. The relationships aren't stable. Markets aren't textbook problems.

The one thing that did translate? Monte Carlo simulation. Not for predicting the future, but for understanding the range of what could happen. That's the whole point.

My Honours Thesis

"Teaching Old Models New Tricks: Comparative Analysis of Classical and Machine Learning Models for FX Forward Contract Evaluation Through Future Spot Rate Prediction Across Developed and Emerging Markets."

Long title. Complex topic. One real finding: ML doesn't automatically beat classical models. Classical models have economic theory behind them. ML just optimizes for the training data. Sometimes classical wins. Sometimes ML wins. It's not about which is better — it's about understanding when each approach makes sense.

I learned something important from that research. ML is a tool, not a magic wand. You need domain expertise to use it properly. Otherwise you're just throwing data at an algorithm and hoping something sticks.

ML Is A Lie (For Most People)

Let me be direct. I tried ML in trading. Hard. And the narrative around it is completely overblown.

Overfitting is the default. Thousands of features, limited data — your model will find patterns that don't exist. Backtests look incredible. Forward tests fail. The model memorized the past instead of learning anything useful.

Data leakage is everywhere. Accidentally include future information in your training set, build a model that seems to predict perfectly, deploy it and lose money. I did this multiple times. Each time I thought I'd found something. Each time I was wrong.

Feature engineering is guesswork. You try hundreds of features, keep the ones that work in backtests, and call it research. That's not science. That's selection bias with extra steps.

You need deep domain expertise to do ML properly. Years of experience understanding how markets behave, what features actually capture edge, how to validate properly. If you're not that person, you're not doing ML. You're doing high-tech gambling with extra steps.

I wasn't that person. Most people who try ML in trading aren't that person either.

Why I Stopped Discretionary Trading

Discretionary trading sounds appealing. You look at a chart, you see a pattern, you make a call. You feel smart. You feel in control.

That feeling is a trap. The market doesn't care how smart you feel.

The problem with discretionary trading is consistency. You make good decisions sometimes and bad decisions sometimes, and you can't tell which is which until the trade is closed. There's no system. No process. Just you, guessing, hoping, and fighting your own emotions.

Every discretionary trader says they can control their emotions. They can't. The moment you see red on your screen, your brain does things you didn't plan for. Exit early. Double down. Revenge trade. You think you're different. You're not.

That's why I moved to systematic trading. Not because I'm smarter. Because I know I'm not smart enough to beat the market with feelings.

Hypothesis Testing Is The Way

Now every strategy I build starts with a hypothesis. Something testable. Something with clear rules. Something that can be proven wrong. If you can't prove it wrong, it's not a hypothesis — it's a hope.

Hypothesis: Markets that are ranging will mean revert within a certain timeframe. Test: Build a system that trades mean reversion in ranging markets. Run it. See if it works.

Hypothesis: Trend following works in strong trends. Test: Build a system that trades with the trend under certain conditions. Run it. See if it makes money.

If the data says your hypothesis is wrong, you don't double down. You don't tweak parameters until it works. You accept it and move on. That's the difference between trading and gambling. Gamblers double down when they're wrong. Researchers move on.

The system I run now has regime detection, Monte Carlo simulation for stop placement, Markov chain filtering, Kelly Criterion for position sizing, and hard risk limits. 1% max per trade. 3% daily limit.

Does it work all the time? Nothing works all the time. But it loses less when it's wrong. It survives drawdowns. And it's testable — I can run the same tests again and validate the results.

The Brutal Truth

Algorithmic trading won't make me invincible. It won't make me rich yet. It won't make the market stop doing what it does.

But it makes me consistent. It makes me survive longer. And most importantly, it makes me lose less when I'm wrong. That's the actuarial promise — not that you won't lose, but that you'll lose less than you would have otherwise. Over time, that's everything.

The code is private. Not because it's special, but because it's not finished. It lives on my server somewhere, running, learning, losing less every day.

I'm in it for the long haul.