Iran war live: IRGC claims new attacks on US warships over naval blockade

The Hook

Today: a live military escalation in the Strait of Hormuz is moving energy markets in real time, institutional money is quietly rotating out of the world's largest ETF complex, and a rigorous ML library just solved a long-standing feature-selection problem for messy, real-world tabular data. The substance is below. It takes about four minutes.

The Cold Open

The Strait of Hormuz at its narrowest point is a corridor through which a substantial share of the world's oil supply moves every single day. This morning, that corridor became the center of a military confrontation that risk desks and energy traders are still trying to fully price. The claims are ahead of the confirmations. The market reaction is not waiting. On mornings like this, the gap between fast, structured intelligence and a browser full of open news tabs is measured in real dollars. Welcome to the show.

The Signal

Iran war live: IRGC claims missile strikes on US warships

The Islamic Revolutionary Guard Corps announced Saturday that it fired ballistic missiles at a US aircraft carrier and a destroyer operating in the Strait of Hormuz, escalating a standoff tied to an ongoing naval blockade. The claim is unverified by US military sources as of publication, but the market impact is not waiting on confirmation: crude futures moved sharply upward, tanker insurance premiums are spiking, and supply-chain risk desks globally are re-running scenarios in real time. The Strait of Hormuz is not a regional story — a significant share of global oil transits it daily. Any real interference shocks energy prices, freight costs, petrochemical inputs, and the inflation models every major central bank is watching. For practitioners using AI tools for market intelligence, this is a case study in asymmetric information risk: the cost of being slow to integrate a genuine escalation signal is far higher than the cost of a false positive. Gemini 1.5 Pro with Google Search grounding can surface corroborating signals across wire services in near-real-time — worth keeping open on a morning like this one.

Quanex (NX): Execution story in a flat-revenue world

Quanex Building Products reported third-quarter results that do not make headlines but should make operators and CFOs take notes: margin expansion plus $42 million in debt repayment on just 1.3% revenue growth. That is a genuinely difficult thing to do. Most companies growing at 1.3% are treading water on margins; NX used operating discipline — cost structure, working capital management, and likely mix shifts — to widen them anyway while simultaneously deleveraging its balance sheet. The strategic read: in a period of soft housing volumes (their core end market), Quanex is proving the business model works even when topline tailwinds are absent. For AI users, this is the kind of earnings narrative that a well-prompted LLM can surface in seconds — the delta between revenue growth and margin or debt trajectory is often buried in footnotes that analysts skim. A Gemini query against an earnings transcript, framed to surface divergences between topline and operating metrics, can replicate this analysis across an entire sector in one pass.

ETF League Tables: Hefty outflow from iShares

Institutional money moved this week — and the direction was out of iShares, BlackRock's ETF flagship, which logged what ETF.com's league tables flagged as a 'hefty' outflow. The meta-signal is consistent regardless of which specific fund drove it: when the world's largest ETF complex sees notable outflows, something is being repriced — whether that is sector rotation, risk-off positioning, or reallocation toward alternatives including AI infrastructure plays that have been drawing capital. ETF flow data is one of the cleaner real-time institutional signals available to both retail and professional investors. It is now also parseable at scale: tools like Vertex AI data connectors can be pointed at structured financial data feeds to flag flow anomalies before they surface in analyst notes, giving teams with the right infrastructure a meaningful lead on where institutional positioning is shifting.

heteroknockoffpy 0.3.1: Feature selection for real-world data

A new PyPI release that most practitioners have not encountered yet addresses a problem anyone who has worked with mixed-type tabular data knows intimately: how do you reliably identify which features actually matter when your dataset combines continuous measurements, categorical codes, ordinal ratings, and binary flags in the same table? The knockoffs framework generates synthetic variables that mimic the correlation structure of real features but carry zero predictive signal. Any feature your model prefers over its knockoff twin is genuinely informative; everything else is noise. heteroknockoffpy extends this to heterogeneous data using conditional residuals and random forests, meaning it handles the messy column types that characterize real enterprise datasets. This is a patch release on a stable codebase — the methodology is not experimental. Niche reach today, but this is the kind of rigorous feature-selection tool that separates ML practitioners who prove their models from those who assume them.

Quick Hits

• PyTorch's Inductor CI pipeline (ciflow/inductor/196137) tagged a new release — Inductor is PyTorch's compiler backend for deep learning inference optimization; activity here is a proxy for the pace of production inference work at scale.
• Energy and shipping stocks are the immediate read-through from the Hormuz escalation — tanker operators, LNG producers, and energy infrastructure names are most directly exposed in the near term.
• — heteroknockoffpy is its practical extension for the messy mixed-type datasets that real enterprise ML actually runs on.

The Anchor

The Strait of Hormuz: Why every risk model is running right now

The IRGC's claim that it fired ballistic missiles at a US aircraft carrier and destroyer is, as of this writing, unverified by US military sources. That matters for factual precision. It does not change what risk desks, energy traders, and supply-chain operators are doing right now — which is running scenarios.

The Strait of Hormuz is one of the most strategically critical chokepoints on the planet. At its narrowest point, it forms a remarkably tight chokepoint. A significant share of global oil supply — and a notable volume of LNG — transits it daily. Any sustained interference with passage does not merely spike crude prices; it cascades through petrochemicals, fertilizers, shipping insurance, freight rates, and the inflation models that central banks in every major economy use to set monetary policy. This is not a regional story with global implications. It is a global story that happens to be located in a single strait.

The military dimension is layered. The US Fifth Fleet operates in this water with significant defensive and offensive capability. But the IRGC's capacity for asymmetric disruption — fast attack boats, anti-ship missiles, historical mine deployment — is real and well-documented. A standoff at this scale, if it sustains, stress-tests the entire doctrine of freedom of navigation that has underpinned global maritime trade since the postwar era.

For readers using AI tools professionally: this is a live case study in where real-time, grounded AI search creates genuine operational advantage over legacy news-monitoring setups. A team running Gemini 1.5 Pro with Google Search grounding can surface corroborating signals — ship positioning data from open-source maritime tracking, commodity futures movement, wire-service confirmation patterns — in a unified query response faster than any traditional news desk can synthesize. The asymmetry of geopolitical risk makes this exactly the use case where the investment in AI-augmented intelligence workflows pays back immediately.

Watch the next 12 hours. If US sources confirm the attack or if open-source ship tracking shows evidence of damage or diversions, the market impact escalates by an order of magnitude. If the IRGC claim goes uncorroborated, expect crude to give back some of the early move. Either way, the risk premium has already repriced. AI-augmented intelligence teams are the first to know which direction it resolves — and the first to position accordingly.

Deep Dive

How knockoffs actually work — and why heteroknockoffpy closes a real gap

Feature selection is one of the most practically important and least rigorously practiced disciplines in applied ML. Most teams default to permutation importance, SHAP values, or correlation thresholds — each of which has well-documented failure modes when features are correlated, especially with mixed data types. The knockoffs framework offers something those methods cannot: a statistical guarantee on false discovery rate.

The core idea is elegant. For each feature X in your dataset, the knockoff framework generates a synthetic variable X̃ — a 'knockoff' twin — that has exactly the same marginal distribution and correlation structure as X, but is conditionally independent of the outcome Y given X. In plain terms: the knockoff variable looks statistically identical to the real feature from a correlational standpoint, but carries zero predictive signal about what you are trying to predict. You then run your model on both the real features and their knockoff twins, measure feature importances, and identify any real feature that your model consistently prefers over its knockoff twin by a statistically meaningful margin. Those features are genuinely informative. Everything else, by construction, is noise.

The original formulation worked cleanly for continuous Gaussian data under linear model assumptions. A later Model-X extension removed the linear model assumption, enabling knockoffs to work with any predictive model — random forests, gradient boosting, neural networks. But both formulations assumed continuous or at least well-behaved feature distributions.

The gap that heteroknockoffpy closes is heterogeneous data — the actual column types that enterprise ML tables contain. Continuous measurements, categorical codes with no natural ordering, ordinal ratings, binary flags, and count variables all have fundamentally different distributional structures. Building a valid knockoff for a five-level ordinal variable is not the same problem as building one for a continuous float. The library addresses this using conditional residuals and random forests, which can model complex conditional distributions without making parametric assumptions about each column type.

The practical consequence of getting this right: you end up with a pruned feature set where each included variable is genuinely informative, with a controlled false discovery rate — meaning you can bound the proportion of selected features that are noise. The alternative — running SHAP or permutation importance on a mixed-type dataset with correlated features — routinely distributes importance across correlated clusters rather than isolating genuinely causal variables. You include redundant predictors, overfit to training patterns, and explain less variance on holdout data than a properly pruned model would.

Version 0.3.1 is a patch on a stable codebase — not an experimental release. If you are building production ML models on mixed-type tabular data, this is worth adding to your feature engineering evaluation toolkit before your next model iteration.

One Technique

Gemini with grounding for geopolitical risk triage

When a breaking geopolitical or macro event hits — like this morning's Strait of Hormuz escalation — most analysts open news tabs and refresh. A faster and more structured approach: open Google AI Studio, enable Gemini 1.5 Pro with Google Search grounding, and run a structured risk-triage query. The output you want is not a news summary — it is a signal matrix: (1) confirmed versus claimed facts, clearly separated; (2) the commodity and asset classes with direct near-term exposure, named specifically; (3) two or three historical analogs and how those resolved, including the timeline of market impact; and (4) the three to five specific data points you need to watch in the next 12 hours to distinguish escalation from de-escalation. Structuring your query around these four outputs turns Gemini from a summarizer into an intelligence-layer tool. The grounding ensures you are not getting training-data responses on a live event — you are getting real-time synthesis from current sources.

One Prompt

Use this in Google AI Studio (Gemini 1.5 Pro, Google Search grounding enabled) when a breaking geopolitical or macro event hits:

You are a senior risk intelligence analyst. A breaking event is developing: [describe the event in one sentence].

Give me a structured triage in four sections:
1. VERIFIED vs. CLAIMED: Separate confirmed facts from unverified claims as of right now. Be explicit about sourcing.
2. EXPOSURE MAP: List the asset classes, commodities, sectors, and geographies with direct near-term exposure. Be specific — name the instruments, not just the categories.
3. HISTORICAL ANALOGS: Name 2-3 comparable past events, how they resolved, and the timeline of market impact in each case.
4. WATCH VARIABLES: List the 3-5 specific data points or confirmations I should monitor in the next 12 hours to determine whether this escalates or de-escalates.

Be concise. Flag uncertainty explicitly. Do not summarize the news — give me the intelligence layer.

Replace [describe the event in one sentence] with today's situation. The four-section structure forces Gemini to separate fact from claim — the most critical discipline in fast-moving events — and turns the response into a decision-support tool rather than a news digest.

One Tip

Turn on Google Search grounding by default in AI Studio

If you use Google AI Studio for any analysis touching current events, market data, or recent product releases — and you are not enabling Google Search grounding — you are getting a model reasoning from its training cutoff, not from today. The toggle lives in the right-hand panel under 'Tools.' Turn it on. Set it as your default for any research-oriented session. The quality uplift on time-sensitive queries is significant; the cost uplift is minimal. This single habit separates AI Studio as a live intelligence tool from AI Studio as sophisticated autocomplete. On a morning when a military escalation is moving commodity markets, the difference between grounded and ungrounded is the difference between a current synthesis and a history lesson.

Tool of the Day

Google AI Studio — Gemini 1.5 Pro with Google Search grounding

Today's stories make a specific case for AI Studio, not a generic one. The grounded search mode is what separates it on a morning when live events are moving markets and unverified claims need to be triaged against real-time sources. Genuinely good for: structured intelligence queries on breaking events, earnings transcript analysis with follow-up questions, and multi-document synthesis where you need to track provenance. Honest limits: it is a research and prototyping environment, not a production pipeline — for production workflows, Vertex AI with the same model family is the right path. The free tier of AI Studio is generous enough that most individual practitioners can do serious research work without hitting token ceilings. If you have not opened it yet this morning, today is a good day to start.

Signature Bites

• The strait matters more than the claim: Whether the IRGC's missile announcement holds up or not, 17 million barrels per day just got a risk premium — and that does not fully unwind on a denial.
• 1.3% revenue, widening margins: Quanex's Q3 is a masterclass in what operational discipline looks like when topline tailwinds are absent.
• Knockoffs ≠ heuristics: Controlled false discovery rate is categorically different from SHAP or permutation importance. Most practitioners have not made this upgrade yet.
• ETF outflows are a leading indicator: When the world's largest ETF complex sees notable outflows, something is being repriced — the question is always which direction, not whether to notice.

Joke of the Day

I asked Gemini to triage geopolitical risk in the Strait of Hormuz. It gave me a four-section structured intelligence brief — verified versus claimed, exposure map, historical analogs, watch variables. Perfect. Then I asked it to help me plan a weekend barbecue. It said: 'I would recommend grounding that query with a more reliable data source.'

Fact of the Day

The knockoffs statistical framework was developed to provide exact FDR control in finite samples for linear models, without requiring knowledge of the joint distribution of features. A later Model-X extension removed the linear model assumption entirely, enabling knockoffs to wrap any predictive model including random forests and neural networks. That 2018 paper is the direct theoretical ancestor of heteroknockoffpy's approach to mixed-type data.

Stat That Matters

The volume of crude oil and petroleum products transiting the Strait of Hormuz daily represents a significant share of global consumption. This single number explains why a military standoff in a 21-mile-wide waterway moves energy markets, petrochemical inputs, fertilizer prices, freight rates, and inflation models simultaneously. It also explains why AI-augmented intelligence workflows — the kind that can synthesize corroborating signals across dozens of sources in seconds — have genuine operational value on mornings like this one.

Trends

Funding leads today's corpus — capital is still moving aggressively into AI infrastructure despite macro uncertainty, and the Hormuz escalation will test whether that conviction holds through a genuine risk-off event. Agentic AI follows as the second-largest category, suggesting the market has moved past the 'will agents work' question and into 'which agents at what scale and at what cost.' Security and policy also contributed meaningfully to today's coverage — regulatory surface is hardening in parallel with capability growth, not lagging behind it as it did in earlier technology cycles. underscores how compressed the intelligence cycle has become. What once required a week of industry monitoring now arrives before noon.

Bold Prediction

If the IRGC's ballistic missile claim is corroborated by US military sources or open-source maritime tracking within 24 hours, Brent crude will trade above $95 per barrel by next week's close — a level not seen since late 2023 — and at least one major shipping insurer will announce a temporary suspension of new tanker coverage for Hormuz-transiting vessels within 48 hours. The catalyst is not the missiles per se; it is the demonstrated willingness to target US naval assets directly, which changes the insurance and operational risk calculus for every commercial ship operator globally, independent of what any government says publicly about the incident.

Paper Watch

This is the foundational paper that generalized knockoffs beyond linear models to arbitrary predictive models, including random forests and neural networks. The key theoretical result: you only need to know the marginal distribution of your features — not the conditional distribution of Y given X — to construct valid knockoffs and control FDR. This removed the primary practical barrier to applying knockoffs to real machine learning pipelines. If you build ML models on real-world tabular data and care about whether your selected features are genuinely informative rather than correlational artifacts, this is one of the most useful 30-page reads in applied statistics from the past decade.

Founder Spotlight

The builder worth watching today is the team behind heteroknockoffpy. Publishing a rigorous statistical feature-selection library to PyPI — without marketing apparatus, without a launch campaign, without a demo video — is the signal of practitioner-first development that produces genuinely useful infrastructure. Version 0.3.1 on a stable codebase indicates this is a maintained tool, not an academic prototype looking for a citation. In a market crowded with AutoML wrappers, dashboard-first ML platforms, and LLM-adjacent frameworks, someone building principled statistics tooling for practitioners is doing the compounding work. The strategic move worth watching: if this library gets adopted by one major ML platform — Vertex AI feature engineering pipelines, a popular feature store, a major data science notebook environment — it moves from niche to standard almost overnight. That is how infrastructure tools scale in the ML ecosystem.

Quote

'Can Execution Outrun Soft Volumes?'

— Insider Monkey headline on Quanex Q3 results. It is the right question — and not just for window and door manufacturers. The companies that are building an affirmative answer into their financials right now — margin discipline, debt reduction, working capital efficiency — are the ones that emerge from a soft-demand period in a structurally stronger position than when they entered it. That gap between those who execute and those who wait for volumes to return is often where durable competitive advantage is built.

Learner's Edge

False Discovery Rate control in machine learning feature selection

When you select features — asking which of your 50 variables actually matter — you are simultaneously running dozens of statistical tests. The classical problem: the more tests you run, the more likely you are to find false positives purely by chance. At a significance threshold of p < 0.05 across 50 features, you expect roughly 2-3 false discoveries even if none of the features are genuinely informative. The standard fix — Bonferroni correction — controls the probability of any single false positive, but at the cost of statistical power. You miss real signals to avoid false ones.

False Discovery Rate (FDR) control takes a different approach: it bounds the expected proportion of selected features that are false discoveries, rather than bounding the probability of any individual error. This gives you a principled way to say 'at most 10 percent of my selected features are noise' — and to actually trust that bound. The knockoffs method achieves FDR control with a finite-sample guarantee, meaning it holds even with moderate sample sizes, not just asymptotically. Most feature-selection methods — SHAP, permutation importance, correlation thresholds — provide no such guarantee. That gap is what rigorous practitioners close.

Sign-off

That is the Gemini Signal edition for September 6th. The next 12 hours in the Strait of Hormuz will determine whether today's risk repricing holds or unwinds — worth keeping a Gemini grounded query open as confirmations or denials come in. See you tomorrow.

Sources

1. Iran war live: IRGC claims new attacks on US warships over naval blockade — aljazeera.com
2. Quanex (NX) Expanded 3Q Margins and Repaid $42M of Debt on 1.3% Sales Growth. Can Execution Outrun Soft Volumes? — Insider Monkey
3. ETF League Tables: Hefty Outflow from iShares — etf.com
4. heteroknockoffpy 0.3.1 — pypi.org
5. ciflow/inductor/196137 — github.com