<?xml version="1.0" encoding="UTF-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <title>UZU News</title>
  <subtitle>UZU News publishes independent coverage of AI forecasting and financial modeling.</subtitle>
  <link href="https://uzunews.com" rel="alternate" type="text/html" />
  <link href="https://uzunews.com/atom.xml" rel="self" type="application/atom+xml" />
  <id>https://uzunews.com/</id>
  <updated>2026-08-25T19:04:54.392Z</updated>
  <entry>
    <title>A high backtest Sharpe ratio proves little: what real validation requires</title>
    <link href="https://uzunews.com/investing/why-a-great-backtest-proves-little-3.html" rel="alternate" type="text/html" />
    <id>https://uzunews.com/investing/why-a-great-backtest-proves-little-3.html</id>
    <summary><![CDATA[Reported Sharpe ratios inflate with every unreported trial. The Deflated Sharpe Ratio and SR 11-7 set out what valid backtest evidence actually requires.]]></summary>
    <published>2026-08-25T08:58:09.000Z</published>
    <updated>2026-08-25T08:58:09.000Z</updated>
    <author>
      <name>Karim Al-Rashid</name>
    </author>
  </entry>
  <entry>
    <title>What does out-of-sample mean in a forecasting paper?</title>
    <link href="https://uzunews.com/analysis/what-does-out-of-sample-mean-forecasting-3.html" rel="alternate" type="text/html" />
    <id>https://uzunews.com/analysis/what-does-out-of-sample-mean-forecasting-3.html</id>
    <summary><![CDATA[Out-of-sample scoring withholds data from every stage of fitting. Here is how the split is done, why it is still overstated, and what it can never establish.]]></summary>
    <published>2026-08-25T08:58:08.000Z</published>
    <updated>2026-08-25T08:58:08.000Z</updated>
    <author>
      <name>Sofia Lindqvist</name>
    </author>
  </entry>
  <entry>
    <title>SEC settles its first AI-washing cases for a combined $400,000</title>
    <link href="https://uzunews.com/finance-news/sec-first-ai-washing-settlements-model-claims-3.html" rel="alternate" type="text/html" />
    <id>https://uzunews.com/finance-news/sec-first-ai-washing-settlements-model-claims-3.html</id>
    <summary><![CDATA[Delphia and Global Predictions paid $400,000 combined in March 2024 for overstating machine-learning use. The orders turn on the gap between claimed and deployed models.]]></summary>
    <published>2026-08-25T08:58:07.000Z</published>
    <updated>2026-08-25T08:58:07.000Z</updated>
    <author>
      <name>Naomi Bergman</name>
    </author>
  </entry>
  <entry>
    <title>How many signals were tested before this one worked?</title>
    <link href="https://uzunews.com/investing/how-many-signals-were-tested-before-this-one-worked-b329667a.html" rel="alternate" type="text/html" />
    <id>https://uzunews.com/investing/how-many-signals-were-tested-before-this-one-worked-b329667a.html</id>
    <summary><![CDATA[Multiple testing turns luck into apparent skill. What the correction procedures control, the hurdles the replication literature settled on, and what a corrected t-statistic still cannot tell you.]]></summary>
    <published>2026-08-25T08:52:06.000Z</published>
    <updated>2026-08-25T08:52:06.000Z</updated>
    <author>
      <name>Karim Al-Rashid</name>
    </author>
  </entry>
  <entry>
    <title>What replaced the guidance that has governed bank models since 2011?</title>
    <link href="https://uzunews.com/analysis/what-replaced-the-guidance-that-has-governed-bank-models-since-2011-94446759.html" rel="alternate" type="text/html" />
    <id>https://uzunews.com/analysis/what-replaced-the-guidance-that-has-governed-bank-models-since-2011-94446759.html</id>
    <summary><![CDATA[Federal regulators rewrote model risk management guidance in April 2026, keeping the same validation core built around conceptual soundness and outcomes analysis while making the rules explicitly non-enforceable and leaving generative AI models outside their scope.]]></summary>
    <published>2026-08-25T08:52:05.000Z</published>
    <updated>2026-08-25T08:52:05.000Z</updated>
    <author>
      <name>Sofia Lindqvist</name>
    </author>
  </entry>
  <entry>
    <title>Every backtest is one of many: what the trial count does to measured performance</title>
    <link href="https://uzunews.com/finance-news/every-backtest-is-one-of-many-what-the-trial-count-does-to-measured-performance-bdca43b3.html" rel="alternate" type="text/html" />
    <id>https://uzunews.com/finance-news/every-backtest-is-one-of-many-what-the-trial-count-does-to-measured-performance-bdca43b3.html</id>
    <summary><![CDATA[Ten trials are enough to yield an in-sample Sharpe ratio of 1.57 from strategies with zero expected out-of-sample performance. What validation, multiple-testing corrections and US supervisory guidance actually require of a backtested claim.]]></summary>
    <published>2026-08-25T08:52:04.000Z</published>
    <updated>2026-08-25T08:52:04.000Z</updated>
    <author>
      <name>Naomi Bergman</name>
    </author>
  </entry>
  <entry>
    <title>What is &apos;probability of backtest overfitting,&apos; and how is it actually measured?</title>
    <link href="https://uzunews.com/markets/what-is-probability-of-backtest-overfitting-and-how-is-it-actually-measured.html" rel="alternate" type="text/html" />
    <id>https://uzunews.com/markets/what-is-probability-of-backtest-overfitting-and-how-is-it-actually-measured.html</id>
    <summary><![CDATA[A diagnostic built to answer one question a good Sharpe ratio can't: how much of a backtested strategy's performance is signal, and how much is the number of times it was tried.]]></summary>
    <published>2026-08-25T08:52:03.000Z</published>
    <updated>2026-08-25T08:52:03.000Z</updated>
    <author>
      <name>Naomi Bergman</name>
    </author>
  </entry>
  <entry>
    <title>The Backtest Trap: Why a Great Historical Track Record Doesn&apos;t Prove a Forecasting Model Works</title>
    <link href="https://uzunews.com/economy/the-backtest-trap-why-a-great-historical-track-record-doesn-t-prove-a-forecasting-model-works-e62c198f.html" rel="alternate" type="text/html" />
    <id>https://uzunews.com/economy/the-backtest-trap-why-a-great-historical-track-record-doesn-t-prove-a-forecasting-model-works-e62c198f.html</id>
    <summary><![CDATA[AI-driven trading and forecasting tools are often sold on the strength of impressive backtested results. Here is the mechanism by which those results can be manufactured by chance alone, and why regulators now require disclosures around them.]]></summary>
    <published>2026-08-25T08:52:02.000Z</published>
    <updated>2026-08-25T08:52:02.000Z</updated>
    <author>
      <name>Rekha Patel</name>
    </author>
  </entry>
  <entry>
    <title>AI-washing is not a model failure: what the SEC&apos;s enforcement actually measured</title>
    <link href="https://uzunews.com/business-news/ai-washing-is-not-a-model-failure-what-the-sec-s-enforcement-actually-measured.html" rel="alternate" type="text/html" />
    <id>https://uzunews.com/business-news/ai-washing-is-not-a-model-failure-what-the-sec-s-enforcement-actually-measured.html</id>
    <summary><![CDATA[The SEC's first AI-washing settlements penalized false marketing about artificial intelligence, not a model's forecasting accuracy — a distinction that matters for anyone judging what a validated model requires.]]></summary>
    <published>2026-08-25T08:52:01.000Z</published>
    <updated>2026-08-25T08:52:01.000Z</updated>
    <author>
      <name>Karim Al-Rashid</name>
    </author>
  </entry>
</feed>