Gio Rendel Masagca Rivadillo / 个人资料

Some thoughts I've been chewing on lately around retail algo trading, backtesting, and where machine learning actually fits into all of it. Sharing here as more of a personal reflection than a discussion prompt.

Been sitting on this one for a while. The more quant finance literature I get through, the more convinced I am that retail algo trading is optimizing for the wrong objective entirely.

Every week on here it's the same pattern. Someone posts a Strategy Tester report flexing a 95% win rate or some absurd profit factor, and nobody's asking the question that actually matters: does this thing exhibit any statistical persistence out of sample, or is it just a curve fit artifact of one particular price history?

I'll take a lower CAGR any day if it comes with a stable Sharpe, low parameter sensitivity, and some walk forward evidence that the underlying market microstructure it's exploiting hasn't already decayed. Not overfitting in the classic sense, just regime decay: the statistical relationship the strategy was built on no longer holds, and no amount of re-optimizing the inputs is going to bring it back.

And that's really the part that gets glossed over. A strategy can pass every in sample metric you throw at it and still be dead the moment it goes live, because the thing it was exploiting (a volatility clustering effect, a session based liquidity gap, a correlation between two instruments) simply isn't there anymore. The backtest doesn't know that. It just reports the numbers from a world that no longer exists.

On the Machine learning front, I've got mixed feelings. Overrated, in the sense that half the "AI EAs" floating around are just an LSTM or XGBoost bolted onto raw OHLCV, which isn't machine learning, it's an overfit function approximator with a marketing budget. Underrated, because almost nobody here is touching unsupervised learning, regime detection, clustering, hidden Markov models, anything that isn't just a more computationally expensive way of reinventing an indicator.

Which is kind of the uncomfortable question, isn't it. If your feature set is still RSI, MACD, ATR, Bollinger Bands, are you doing machine learning, or are you spending a lot of compute to rediscover heuristics that have been public domain since the 90s?

There's a version of this that's even more uncomfortable: a lot of these models aren't even failing quietly. They pass cross validation, they pass walk forward, they look statistically sound by every metric people know how to check, and they still collapse in live conditions, because the validation itself was built on the same underlying assumption of stationarity that the strategy is exploiting. If the regime shifts, your validation shifts with it. You don't get a warning.

Honestly, I'd argue most EAs aren't exploiting inefficiencies at all. They're exploiting historical coincidence. Backtesting rewards complexity generously and punishes it almost never.

That asymmetry is worth sitting with for a second. Every added parameter, every extra filter, every additional condition on entry gives the optimizer one more degree of freedom to fit noise. And the Strategy Tester will happily reward that with a smoother equity curve, because smoother in sample is exactly what you'd expect from a system with more knobs to turn. Robustness doesn't show up in that number. It only shows up later, when the market stops cooperating with the specific shape of the noise the system was fit to.

Still working through where I think the next real edge comes from, better validation methodology, more deliberate feature engineering, alternative data, or whether retail is already past the point of meaningfully competing with institutional quant research. No firm answer yet, just where my head's been at lately.

已在MQL5.community注册