Honest experiments
Quantitative strategies and honest backtesting: what we tested and didn't work, with the numbers. Publishing the null results is part of the brand.
We are closing the AI tournament: none beat chance in 315 predictions
After 315 resolved predictions per model, no AI forecast the direction of the market better than chance, and all of them sized the move worse than a rule with no AI at all. We are closing the tournament, and the agent that ranked the Núcleo S&P 500 portfolio too.
Read →Tactical Defensive crosses below the benchmark
It gained 0.40% on the week and still went from ahead of the yardstick to behind it: 3.16% against the equal-weight market's 3.90%. Momentum + Core lost 4.33% in seven days.
Read →All three published strategies are ahead of the market, and the Núcleo holds up across all 21 start dates
As of 9 October, Núcleo S&P 500 (+5.43%), Cesta Diversificada (+5.03%) and Defensivo Táctico (+3.06%) are ahead of the market (+2.73%) in simulation. And the Núcleo backtest, repeated on each of its 21 possible start days since 2022, beats the market on all 21.
Read →AI tournament, five weeks in: none beats plain volatility, and an error of ours skewed the A/B
Five language models have spent five weeks forecasting stocks in public and none beats a mechanical volatility forecast. And the test saying that giving them the earnings context made everything worse was measuring an error of ours: the announcement dates came from misclassified news, and none of 35 was a real report.
Read →Núcleo S&P 500 falls below the benchmark
The bot went from 7.43% to 4.80% since seeding while the house yardstick slipped from 7.21% to 6.71%. It is the only crossing this week, and it is downward.
Read →Momentum + Core at -14.52% and demoted to experimental; the house benchmark sits at +7.1%
Six of seven simulated portfolios fell this week. The worst is down 14.52% since seeding and has lost its public label. The equal-weighted reference panel is up 7.1% over the common window.
Read →We froze OKX's top 50 and watched for seven weeks. The scoreboard says +1.20%; the actual return, −75 basis points
In July we wrote down which test we would run and when. Today was the day. With the cohort frozen in advance and 4,142 positions captured hour by hour, persistence of skill hasn't just weakened — it has vanished. And the market was up 21.6% while they lost money.
Read →We measured seven kinds of corporate events. Only three move the stock, and 42% of the 'news' is noise
We classified 7,046 news items into seven event types and measured whether they were followed by more movement than usual — not by rises or falls. Earnings, executive changes and guidance, yes; analyst ratings, capital operations, regulation and M&A, no. Ten years of SEC filings confirm the pattern. And the criterion we fixed for adopting a filter failed, so we are telling you.
Read →We spent a year measuring with an instrument six times too blunt
We found out that the way we judged our signals was throwing away 95% of the data. We re-measured ten signals with the right method — momentum and low volatility included — and here is what came out.
Read →48 days of live forward testing: the full scoreboard, including the strategy that is down 10.6%
We seeded eleven strategies in paper on June 18th and committed to publishing the result win or lose. Here it is in full: the one beating the market (and how much of that edge evaporates against the right benchmark), the one collapsing, the one we already called a mirage — and why after 48 days none of it means anything yet.
Read →We put OKX's 50 most-copied traders under the microscope: the showcase shines, persistence says noise
6,496 exchange-verified trades over 3 months, serious statistics: 9 of 24 traders look like geniuses — until the persistence test. Here's how a copy-trading leaderboard manufactures its shine.
Read →The truth about Binance grid bots
We backtested the most popular grid bots on real history. They win sideways, sink in the crash, and the 7-day drawdown hides the risk. The numbers, unvarnished.
Read →Now we validate grid bots from 7 exchanges, not just Binance
We expanded the free tool: paste any grid bot from Binance, Bybit, OKX, KuCoin, Bitget, Gate.io or MEXC and we'll tell you the truth on real history.
Read →Paste any Binance bot and we'll tell you the truth
New free tool: paste any Binance grid bot's URL and we backtest it on the fly over real history, no make-up.
Read →What we have today (and what we don't): an honest status
No smoke: what's built and working, what's in forward validation, and what doesn't exist yet.
Read →Why your backtest lies: survivorship bias
If you test your strategy only on the companies still in the index today, you measure a return that never existed.
Read →We tested news sentiment. It doesn't predict the market.
Catalyst-news sentiment measured live: indistinguishable from zero (and, with cost, negative). The numbers, unvarnished.
Read →Cost: the wall that knocks down most short-term signals
Many strategies win gross and lose net. Over short horizons, fees and spread eat the edge.
Read →What an honest verdict is: why a high Sharpe isn't enough
Date-clustered bootstrap, PSR and percentage of positive years. How we decide if a strategy truly beats chance.
Read →Overfitting: how a thousand variants fool you without noticing
Try enough combinations and one will look like gold by pure chance. Here's how we control overfitting.
Read →Horizon matters: why the edge lives at a month, not intraday
We tested from 1 minute to several months. The sweet spot with real edge is 1–3 months, not the short term.
Read →Stock ranking vs. market timing: what really works
Guessing when the market rises or falls came out null again and again. Ranking which stocks to hold did not.
Read →When long-only beats long/short (and why)
Shorting doesn't always add: it has borrow cost and, in certain regimes, losing on the short leg destroys the result.
Read →Market regime: sometimes the best trade is to sit out
A regime filter that keeps you in cash at the wrong moment can be worth more than any selection signal.
Read →Academic backing vs. promises: telling signal from marketing
A credible edge usually has peer-reviewed literature behind it. If it only has a pretty chart, be skeptical.
Read →