Two ML Models, Same Universe, Completely Different Personalities

Most AI and algo trading platforms are a black box.

You pick a strategy. It runs. You hope it works. You never know why it worked or why it stopped.

That's not how machine learning works. With ML, you don't program in the rules. You give the model a list of possible signals and let it figure out which ones actually mattered historically. You provide the ingredients. It finds the recipe.

I have two models that make this concrete.

Same Setup, Different Outcome

Both models run on the QB500 universe. Same timeframe: 3 years. Same target: +2.5% in 4 days. The only difference is I tweaked some of the features I gave each one and retrained.

They're iterations of each other. They ended up with completely different personalities.

What Model 1 Learned

The top feature Model 1 landed on was the 2-10 Year Yield Spread — at 10.0% of total weight. That's a macro signal. It reflects whether the bond market thinks the economy is healthy or heading for trouble.

The rest of its weights are spread across 40+ features fairly evenly. No single signal dominates. It's running a balanced committee vote every day — multiple inputs, distributed influence, no single thing it's betting everything on.

Feature mix: 54% individual stock signals, 46% broad market and macro.

What that looks like in practice: Model 1 made 10–20 picks every single day from late March through April. Didn't matter what the market was doing. Consistent participation. Steady.

Live stats: 61.6% avg confidence. +2.86% avg return. 89% win rate.

It grinds. It shows up every day, takes its swings, and grinds higher.

What Model 3 Learned

Model 3 landed on a completely different worldview.

Its top three features: QB1000 P/E Median at 16.6%, Market Trend Strength at 14.6%, Nasdaq-100 RSI at 11.6%. Three features doing most of the heavy lifting. And two of them are explicitly about whether the market is trending and whether valuations support that trend.

This model is not interested in participating unless the big picture is confirmed.

Feature mix: 53% individual stock signals, 47% macro. Similar split — but concentrated where Model 1 was balanced.

What that looks like in practice: during March 2026, when markets sold off on Iran war fears, Model 3 went nearly silent. A handful of picks. In the first week of April, as the market started turning, still very few picks. The trend hadn't confirmed yet.

Then the move held.

Model 3 went all in. Over 100 picks in a single day. The market continued higher after that.

Live stats: 58.5% avg confidence. +3.86% avg return. 85% win rate.

It waits. It watches. And when everything it cares about lines up, it bets big.

The March and April Story

The market context matters here, because it shows how differently these two models responded to the same environment.

March 2026: Geopolitical shock. Markets sold off hard. Rough period across the board.

Model 1 — the balanced macro watcher — pulled back significantly on pick count. Its top feature, the 2-10 yield spread, moves slowly. When that spread signaled caution, Model 1 listened and went quiet. When the market started recovering in April and the spread confirmed it, Model 1 ramped up gradually — a handful of picks at first, then building as the trend held.

Model 3 — the trend and valuation watcher — saw an unconfirmed trend and stayed almost entirely out. Zero conviction during the selloff. But when the market turned and held the turn, its confirmation signals all fired at once. It went from near-zero picks to over 100 in a single session.

Both were right. They just expressed it differently.

Neither Is Better

Model 1 is a participant. It shows up every day, takes its swings, and grinds. If you want steady daily exposure and don't want to time the market, this is the personality you want.

Model 3 is a conviction trader. It waits. It misses the early part of moves sometimes. But when it acts, it acts with size and confidence. If you want big concentrated bets when the big picture aligns, this is your model.

Both made money. Both worked over the same 30-day period. They just did it in completely different ways.

The Point

On a rule-based platform, you'd never know any of this. You'd just see a strategy that either worked or didn't. You'd have no idea whether it was working because of one signal or twenty. You'd have no way to intentionally build a different personality into the next version.

With a trained ML model, you can see what it learned. The feature importance chart tells you exactly which signals the model decided mattered — and how much weight it gave each one. Change the feature menu, retrain, and you might end up with a completely different worldview. That's not a bug. That's the process.

The part that surprises most people when they first start iterating: you expect one version to be clearly better. Instead you end up with two different worldviews that both make money in different ways.

You give it the ingredients. It builds the recipe. And then it shows you exactly what it decided to cook.