Our forecasting track record

What changed, in plain words

• Prices: US wholesale inflation was flat in July and 4.7% higher than a year earlier, down from 5.5%. Economists had expected a 0.2% monthly rise, so this came in cooler than forecast. Wholesale prices are what companies charge each other before anything reaches a shop shelf. — The cost pressure behind the prices you pay eased, and markets reacted to that.
• The fine print: almost all of that relief was a 5.7% drop in wholesale petrol prices. Strip out food, fuel and retail markups and the same report's underlying measure sped up to 0.4% in July from 0.1% in June, while the earliest-stage input costs are still 9.7% higher than a year ago. — The good news was mostly cheap fuel; underneath it, material costs firmed.
• Oil: the International Energy Agency now sees the world burning LESS oil in 2026 than 2025 — a first since Covid — and Brent fell 2.1%. The cut is 1.6 million barrels a day, blamed on high fuel prices and shortages caused by the closed Strait of Hormuz, which remains shut. — Cheaper oil here is not the blockage clearing; it is costly fuel cutting how much gets used.
• AI chips: Goldman Sachs estimates a third of this year's $750bn of AI construction spending is funded with borrowed money. Next year's total is put near $1.2 trillion. On the same day, AI chipmaker Cerebras fell about 16% after its sales missed forecasts even as it raised its outlook for the year. — The AI buildout leans on debt more than profits, and investors are less forgiving of misses.
• Chips in your basket: a Bank of America economist noted computer, software and smartphone prices rose 1.4% in one month. His reading is that business demand for AI is raising input costs, which are being passed on to shoppers — a route from data centres to household bills that has not appeared in this record before. — The AI boom has started showing up in the price of ordinary electronics.
• Your bills: the average US 30-year mortgage rate fell to 6.67% from 6.69%, its first decline in six weeks, and traders cut the chance of a Federal Reserve rate rise in September to 35% from about 50% two days earlier. — Borrowing got very slightly cheaper, and the market now thinks a rate rise next month is less likely than it did on Tuesday.
• Jobs: new US unemployment claims rose by 9,000 to 209,000, more than the 202,000 economists expected, but the smoother four-week average held at 199,000 and the number of people still claiming fell by 22,000 to 1.78 million. — One noisier week, no change in the underlying picture — the job market is not obviously cracking.
• Gold: gold touched $4,449 an ounce, its highest since 5 June, before slipping 0.9% to close near $4,366. It is up roughly 10% over the past month. — Gold rising while inflation news improves is a sign some buyers are worried about something other than this month's numbers.
• On the radar: copper — which we monitor but do not hold — sits at $6.57 a pound, down slightly on the day but almost 47% higher than a year ago, with reports of Chinese buyers backing away from the price while mine supply stays tight. — Copper is in everything from wiring to data centres, so a price up nearly half in a year is a cost that eventually reaches other prices.
• Next week: an extra 50% US tariff on certain Canadian dairy, alcohol and motor-vehicle imports takes effect on Wednesday 19 August, and it applies even to goods that normally qualify for the North American trade deal. Home Depot reports results on Tuesday and Walmart on Thursday. — A dated, already-announced tax on some imports arrives next week, and two big retailers will say what they see happening to shoppers.

Headline — the relief was energy arithmetic, and the pipeline underneath it firmed

July PPI final demand was unchanged month-on-month against a +0.2% consensus, and the unadjusted 12-month rate fell to 4.7% from June's 5.5% (BLS, released 2026-08-13; EVT-0373). That was the day's driver: the S&P 500 rose 0.7% to a record close, the Nasdaq added 0.8%, the Dow 0.1%, the 10-year Treasury yield fell to 4.65% from 4.68%, and CME-implied odds of a September Fed hike dropped to 35% from roughly 50% two days earlier (EVT-0374).

The composition matters more than the headline. Final demand goods fell 0.7% on a 3.1% drop in energy and a 5.7% drop in gasoline; final demand less foods, energy and trade services ACCELERATED to +0.4% from +0.1% and is also running at 4.7% year-on-year; stage-1 intermediate demand is +9.7% year-on-year. So the goods-side relief is fuel arithmetic while the underlying pipeline firmed — and the July gasoline print predates the roughly $4.01 August pump prices and the $87–89 Brent that will land in the August CPI released in September. That asymmetry is why the escalated disconfirmer (P-0085) is priced low rather than at a coin flip.

On the energy chain itself, the direction of the news genuinely reversed for the first time in weeks — but for the wrong reason. The IEA cut its 2026 world-demand path to a 1.6 mb/d CONTRACTION, 510 kb/d deeper than July and its first annual-decline call since Covid, attributing it explicitly to high fuel prices and Hormuz-related product shortages; OPEC trimmed 2026 growth to +580 kb/d, a fourth consecutive cut, leaving the two forecasts more than 2 mb/d apart (EVT-0375). Brent fell 2.1% to $87.07. This is demand destruction, not supply normalisation, which is the softest possible way for an energy thesis to be disconfirmed. Policy meanwhile hardened rhetorically without moving materially: Tehran called the strait "under Iran's control and management" against Trump's "total control" and "I THINK WE WILL KEEP IT", with an added Iranian compensation demand over a Qeshm Island oil spill (EVT-0381) — more heat, no new information.

The AI-buildout leg supplied the run's two most substantive findings, both on the financing and pass-through dimensions rather than on demand. Goldman Sachs estimates roughly one third of 2026's ~$750bn AI capex is debt-financed (~$250bn of IG issuance, $194bn already placed in H1), rising to ~35% of ~$1.2trn in 2027 with ~$400bn of direct hyperscaler issuance, and ~$5.8trn of cumulative hyperscaler capex through 2030 with ~$1.3trn of incremental financing need in 2028–30 shifting toward private credit and project finance (EVT-0377). Read against yesterday's Supermicro cash-flow finding (FY26 operating cash flow −$6.81bn funded by $9.48bn of financing inflows, EVT-0368), the picture is consistent: order books are real on the income statement and increasingly borrowed on the cash-flow statement. Cerebras then delivered the market-structure counterpart — GAAP revenue $180m versus ~$194m consensus, an adjusted loss of $0.05 against −$0.17 expected, cloud revenue +281%, and a RAISED full-year core-revenue guide, and the stock still fell ~15.9% to $220.31 (EVT-0378). A guidance raise punished on a revenue miss is the same de-rating tape the memory complex has been trading in.

The genuinely new transmission channel is BofA's decomposition of the July CPI: core goods +0.20% m/m with a 1.4% m/m rise in IT commodities contributing 12 basis points, which Stephen Juneau attributed to business AI demand raising input costs that are passed to consumers (EVT-0376). This is single-source for the attribution and should be treated as an analyst decomposition, not a BLS finding — but it is the first time in this run's record that the AI capex cycle has been named as a live contributor to household goods prices rather than to equity multiples. It also gives P-0058 (August PCE core goods ≥2.0% y/y) a mechanism.

Elsewhere: jobless claims rose 9,000 to 209,000 against a 202,000 median while the four-week average held at 199,000 and continuing claims fell 22,000 to 1,777,000 (EVT-0379) — a noisier week inside an unchanged trend, so the labour thesis neither confirmed nor disconfirmed. Freddie Mac's 30-year fixed fell to 6.67% from 6.69%, the first decline in six weeks, which moves against open claim P-0080 (MORTGAGE30US ≥6.85% by 2026-09-30, p=0.40). Gold printed $4,449.39 intraday, its highest since 5 June, then closed down 0.9% at $4,366.38, up ~10% on the month — gold strengthening into improving inflation data is a discordant note worth watching rather than explaining away. Bitcoin opened at $63,410 with daily, weekly, monthly and yearly trend windows all negative, roughly 50% below its October-2025 peak, on a day equities set a record (EVT-0382), which continues to corroborate thesis #3. On the radar, copper at $6.57/lb is +46.84% year-on-year (EVT-0380, single data provider, flagged).

Calibration note for the Reviewer, recorded rather than acted on: the market's September-hike probability is now 35%, while two open predictions (P-0062 and P-0067) both sit at 0.45 on the same 15–16 September decision. Those standing calls now look high, and the pair remains the duplicate-claims-on-one-event problem the 2026-08-09 review named. No third Fed claim was logged today.

Readings

Fact-check log

Standing theses — re-scored

• #6 Energy & critical-commodity supply — DISCONFIRMING, in the softest possible way. The thesis's own stated disconfirm condition is "demand-driven softness; supply normalises." Today delivered the first half explicitly: the IEA cut 2026 world demand to a 1.6 mb/d contraction and named high fuel prices and Hormuz product shortages as the cause, while Brent fell 2.1% and broke a six-session climb. Supply has NOT normalised — the strait is still shut with no agreement. Recorded honestly as a disconfirm and escalated to P-0085 rather than reframed as a confirm. — The energy thesis took its first real hit in weeks, but because expensive fuel is destroying demand, not because the blockage cleared.
  For your money For a household this is the ambiguous kind of good news: fuel bills ease, which helps what your money buys, but the stated reason — people using less because it costs too much — is also how a slowing economy first appears. An unclear picture is exactly what a cash buffer and savings spread across different things are for, rather than a reason to make a big move.

  What to watch Matters more if the big energy agencies keep trimming how much oil they think the world uses while the route stays closed; less if a reopening is agreed and softer prices are traced to supply returning.
• #1 AI-credit fragility — CONFIRMING on the financing leg, still DISCONFIRMING on the credit leg. Goldman's estimate that ~33% of 2026 AI capex is debt-funded, rising to ~35% of ~$1.2trn in 2027 with the 2028-30 need shifting toward private credit, is direct evidence for the vendor/leverage mechanism, and it corroborates yesterday's Supermicro finding (operating cash flow −$6.81bn funded by $9.48bn of financing inflows). But the thesis's disconfirm condition — "spreads stay tight" — still holds: high-yield spreads are 2.71pp with no widening flag, and investment-grade issuance is being absorbed without visible strain. The mechanism is building; the stress is not yet priced. — The evidence that the AI boom runs on borrowed money keeps accumulating, but lenders are still charging borrowers as little as ever.
  For your money AI-linked companies are now a large share of the main share indexes, so a boom funded by borrowing rather than profits puts fragility inside ordinary index and pension savings. Nothing here is a reason to make a big move — broad diversification, not concentrating in the theme, and not borrowing to invest are the ordinary defences.

  What to watch Matters more if the extra interest riskier companies pay starts widening while this borrowing continues, or if a large AI buildout has to be refinanced on visibly worse terms; less if these buildouts begin funding construction out of the cash they generate.
• #5 Sovereign-debt / rates stress — mildly DISCONFIRMING today, unchanged structurally. The 10-year fell to 4.65% from 4.68% and 4.72% on Monday, and the 30-year mortgage rate fell for the first time in six weeks to 6.67%, both moving against the thesis. But the 30-year yield remains near 5.24% and the 10-year sits far above its 3.97% pre-war level, so the level of the disconfirm is a wobble inside an elevated regime. This also moves against open claim P-0080 (MORTGAGE30US ≥6.85% by 2026-09-30, p=0.40). — Long-term borrowing costs eased slightly, but they remain much higher than before the war, so nothing structural changed.
  For your money Long-term rates set the cost of mortgages and business borrowing while also making safe savings pay decently, so a small fall is a marginal help either way. One week is not a turn — keeping a cash buffer and not stretching on new borrowing remain the sensible postures rather than a reason to make a big move.

  What to watch Matters more if long-term rates climb back toward their highs while government bond sales need higher rates to attract buyers; less if this easing persists for several weeks alongside calmer fuel prices.
• #3 Crypto → AI/compute capital rotation — CONFIRMING again. Bitcoin opened at $63,410 with daily, weekly, monthly AND yearly trend windows all negative — roughly 50% below its October-2025 peak — on the same day US share indexes closed at a record. The speculative sleeve keeps breaking while the AI-levered sleeve keeps making highs. Fifth consecutive session recording this divergence. — Speculative crypto keeps falling while AI-linked shares set records, which is the rotation this thesis predicts.
• #9 Labour-market deterioration — NEUTRAL, neither confirmed nor disconfirmed. Initial claims rose 9,000 to 209,000 and beat the 202,000 median, which points the thesis's way, but the four-week average held exactly flat at 199,000 and continuing claims FELL 22,000 to 1,777,000, which points the other way. One noisier week inside an unchanged trend is not evidence in either direction, and it should not be scored as though it were. — The weekly jobs numbers were mixed and the smoother measures did not budge, so nothing was learned about the job market today.
• #7 Equity concentration / breadth — MIXED. Leadership was narrower by index (Nasdaq +0.8% > S&P +0.7% > Dow +0.1%), which fits the thesis, but the day's biggest sector gains were in real estate and homebuilders (AvalonBay +2.3%, D.R. Horton +2.8%) on the fall in long-term rates — rate-sensitive participation outside the AI complex, which cuts against it. Relevant to open claim P-0070 (equal-weight S&P outperforming by 2026-09-04, p=0.40), which today's rotation modestly helps. — Technology still led, but interest-rate-sensitive parts of the market joined in, so the market's narrowness eased a little.

Research only — not financial advice. The practice portfolio is for learning only; this is a read-only risk briefing meant to help explain risk, never tell anyone what to buy or sell.
Project snapshot — changes waiting for review: 16 · tracked forecasts open: 52 (checking now: 3) · track record so far: 0.118227 · practice portfolio updated: 2026-08-13.
Past briefs & dashboards: market-sentinel-lyart.vercel.app/reports