Any AI can beat the market from 2017-2021. You just have to build it in 2021.
This isn't a bug in quantitative finance — it's a feature of how backtesting works. And it's being used to sell you something.
See how we do it differently
We don't claim to beat the market. We show you the analysis — including where our models disagree.
The Anatomy of a Backtest
Backtesting sounds reasonable: take your strategy, run it against historical data, see how it would have performed. The problem is that when you build the model, you already know what happened.
The data scientist sees 2020. They know which stocks crashed in March and which recovered by December. They can (consciously or not) tune their model until it "discovers" patterns that happened to work during that specific period. This is called overfitting — the model learns the noise of the past, not the signal of the future.
Live performance — where the model makes predictions about data it hasn't seen — is the only honest test. It's also almost always worse than the backtest.
Case Study: Danelfin
Danelfin is a real company, not a scam. Founded by Tomás Diago (who previously founded Softonic), backed by Nauta Capital, with a legitimate product. We want to be clear about that upfront.
Their marketing prominently features claims like:
"Since 2017, our AI Score 10/10 stocks beat the market by +20%."
Here's the thing: Danelfin launched in 2021.
That means four years of that performance (2017-2021) is backtesting — the model "predicting" outcomes that already happened. Only the period from 2021 onward represents actual live predictions. The marketing chart seamlessly blends these two periods with no visual distinction.
When someone shows you returns "since 2017" for a platform that launched in 2021, this is what's actually happening. The first four years are fantasy.
This isn't illegal. It's not even technically dishonest — they're not claiming it was live. But it's designed to create an impression of years of proven performance when the actual track record is much shorter.
What we do differently
We don't backtest. We don't claim to beat the market. Our LLMs read documents and surface patterns in language — they don't predict prices. There's no historical performance to show because we're not making predictions you can score.
The Supporting Cast: Other Red Flags
Backtest inflation rarely travels alone. Here are the patterns that often accompany it:
Survivorship Bias
Signals that went wrong quietly exit the track record. If a "buy" signal turns into a 50% loss, it gets rebalanced out or reclassified. The winners stay; the losers disappear. You only see what survived.
Cherry-Picked Windows
"60-day alpha" gets highlighted because 90-day looks worse. The marketing team found the time window where the numbers look best and featured that one. If they had a better window, they'd show you that instead.
Feature Bloat
"We analyze 10,000 features!" sounds impressive. In practice, more variables often means more overfitting. A simple model with 5 key variables frequently outperforms a model drowning in 10,000 noise signals. This is marketing, not methodology.
Confidence Theater
"87.3% accuracy!" The precise decimal creates an impression of scientific rigor. But accuracy at what? Over what period? Predicting direction or magnitude? The precision distracts from the vagueness of what's actually being measured.
No Drawdown Disclosure
Returns without risk metrics are meaningless. If a strategy beat the market by 20% but had a 60% drawdown along the way, would you have held through it? Returns tell you the destination; drawdowns tell you the journey.
What Danelfin Actually Does Well
We're not here to bury them. Danelfin has legitimate strengths:
- Explainable AI: They use decision trees that show their reasoning — "RSI is low" + "revenue growth is high." This is genuinely better than black-box neural nets that just say "trust me."
- Real product, real funding: Backed by Nauta Capital, founded by an experienced entrepreneur. Not a fly-by-night operation.
- Works as a screener: User consensus (Reddit, reviews) suggests it's useful for surfacing ideas you wouldn't have found otherwise.
What we do differently
We show you when our models disagree. Danelfin shows you confidence scores. One of these helps you understand uncertainty; the other hides it.
How to Evaluate Any AI Platform
Use this checklist when evaluating any platform that claims AI-powered investment insights:
| Question to Ask | Red Flag | What We Do |
|---|---|---|
| When did the platform launch? | Claims before launch date are backtest | We don't claim historical returns |
| Can you see live-only performance? | If they won't separate it, there's a reason | We don't make performance claims |
| Do they show drawdowns? | Returns without worst periods are cherry-picked | We show model disagreements |
| How specific are the claims? | "Beat the market" is vague | We classify, not predict |
| What are users doing with it? | "Trade signals" = dangerous | Research acceleration |
Why This Industry Works This Way
Platforms like Danelfin exist because retail investors want "the answer." They want someone — or something — to tell them what to buy. That's a human need, and it creates a market.
Backtested returns feel like evidence because we're trained to trust charts. A line going up and to the right triggers the same pattern-recognition circuits regardless of whether it represents real predictions or retrofitted fantasy.
The platforms aren't lying — they're just showing you what sells. The interesting question is: what would it look like if they were honest?
What we do differently
We built our platform to show uncertainty, not hide it. When our models disagree, we show you the disagreement. When we don't know something, we say so. This is less marketable than "87.3% accuracy since 2017" — but it's more useful.
The Takeaway
When someone shows you a chart of AI performance since 2017, ask one question: When did the model actually start running?
If the answer is 2021, you're looking at four years of fantasy and three years of reality. The fantasy part always looks better.
We don't know if Danelfin's live performance is good or bad — we haven't tracked it long enough. What we do know is that their marketing makes it very hard to tell. And that's the point.
See how we approach analysis
Explore our equity analyses to see how we show uncertainty, model disagreements, and cross-lens conflicts.
Further Reading
How Frameworks Can Hide Hallucinations
Structure creates false confidence. Tables make numbers look verified when they're not.
A Taxonomy of LLM Hallucinations
Not all errors are created equal. Some are dangerous, some obvious, some almost impossible to catch.
Why We Use Multiple LLMs
Why we show disagreements instead of hiding them behind confidence scores.