Predictions on the Record What the engine said in March. What actually happened. What it says now.

On March 4, 2026 the AI-Stocks engine put a specific 5-year forecast on paper for 30 industries. Five months into a real market, 72.7% of those directional calls have been correct and 81.6% of the money moved the way the engine said. This page is the record, preserved. It is also the next set of forecasts, published on the same terms.

By Scott Covert · Published 2026-07-03 · AI Stock Market Impacts

The most useful thing anyone can do with a forecasting engine is publish the forecasts, preserve the files, and let time grade the work. That is what separates a track record from a marketing claim. The version of this engine that ran on March 4, 2026 produced a full matrix of 1-year, 2-year, 3-year, 5-year, and 10-year projections for every industry it covers. That matrix is preserved as scenarios-computed_2026-03-04_v21-or-earlier_first-published-forecast-anchor.json, and every number on this page traces back to it.

Five months is not five years. The engine was designed to say useful things about the 2-to-5 year window, and the H1 2026 tape is a very short slice. But five months of live data is enough to grade the direction of every call. The result is at the top of the page for anyone who wants to skip to the bottom line.

The forecast the engine put on paper five months ago

The engine on March 4 was 24 versions younger than the current one and used a base scenario of "current energy trajectory, current AI progress rate." It ran across 28 industry groups (two additional splits landed later, taking the count to 30). Every industry got a multiplier at each of five timeframes, where 1.00 means "flat versus a global equity baseline" and, for example, 2.73 means "worth 2.73x its March 2026 relative starting point at the 5-year horizon."

Here are the top and bottom of the March 4 base case at the 5-year horizon:

The March 4 file contained the same shape at every other timeframe. Software Cloud, Biotech-AI, and Chip Design were the biggest structural winners at every horizon. Utilities, communication and professional services, and media were the biggest structural losers. Everything in between compressed toward 1.00.

Two calls in the March file deserve to be named individually because they were unusual for the moment. First, the engine had Utilities as its single worst 10-year industry at 0.28. That was against a consensus in early 2026 that treated utilities as the direct beneficiary of AI datacenter buildout. The engine saw the datacenter capex as flowing to hyperscalers and behind-the-meter dedicated generation, not to public regulated utilities, and it saw regulatory friction on rate cases as capping any upside. Second, the engine had Communication & Professional Services at 0.57 at ten years, worse than anything except Utilities. That was the engine's read that consulting, staffing, ad agencies, and marketing services are the most AI-disruptable service sectors in the economy, because their whole job is language work.

What actually happened, February to July 2026

Every industry in the forecast was mapped to the cleanest tradeable US sector ETF proxy. Where none existed — for example enterprise SaaS AI-native, which is mostly private companies like Anthropic, Glean, Sierra, and Harvey — the industry was excluded from grading. Twenty-two industries got clean grades. Sixteen of them moved the way the engine said. That is a 72.7% sign-match. Weighted by the size of the actual moves, the engine got 81.6% of the money right.

The list of hits keeps going. Biotech-AI (XBI +24.44%): called bullish, correct. Banks (KBE +13.65%): called bullish, correct. Big Pharma (IHE +13.82%): called bullish, correct. Capital Goods (XLI +19.32%): called bullish, correct. Comm Services (XLC -9.79%): called bearish, correct. Energy (XLE +22.63%): called bullish, correct. Tech Hardware (XLK +28.51%): called bullish, correct. Transportation (IYT +16.90%): called bullish, correct.

The six misses and what they mean

Six directional calls went the other way. Named in order of how much they mattered:

The pattern is the important part. Five of the six misses share the same shape. The engine saw the 5-year structural picture correctly and did not downweight the near-term cycle. That is not a random error. It is the specific weakness of a structural model in a narrative-dominated tape, and it is the thing the v25 and v26 engine additions were built to address — new fields for cycle phase, priced-in narrative, customer health, catalyst proximity, capex absorption risk, and the split between consumer resistance and labor resistance to AI adoption. Those additions were shipped on July 2, 2026, one day before this page went up. Their effect on the current forecast is Panel 3.

A note on the null hypothesis. A "the tape is up in a bull market" model would also have hit roughly 73% on H1 2026 (about 22 of 30 sector ETFs finished the half positive). The engine's edge over that naive baseline is not in the raw hit rate. It is in going maximum-conviction bearish on the two negative blocks and having both fall hard — enterprise SaaS legacy at -19.79% and comm services at -9.79% — and in going maximum-conviction bullish on the semi stack, biotech-AI, and banks, all of which were among the biggest movers of the half. A coin-flip model gets the direction right in an up tape. A working engine gets the size of the move right.

The July 3, 2026 forecast, in two scenarios

The current engine has been through 24 additional calibration rounds since March. It carries 20 fields per industry — up from 11 in March — including cycle phase, priced-in narrative, capex absorption risk, and split resistance scores for consumer and labor separately. It runs on 30 industry groups instead of 28. The full field lineage is preserved in the changelog. The two scenarios below both run on current-trajectory energy (roughly today's grid, no fusion breakthrough, no rolling blackouts). The base case assumes AI capability improves at roughly today's pace. The Faster case assumes it accelerates.

Base case, current trajectory — top and bottom of the 30-industry matrix at 5 years

Two things about the current base case are worth naming directly. First, the top of the matrix is now shorter and steeper than it was in March. The winners have less headroom because more of the story is priced in — v25 and v26 added a pricedIn field and an aiMaterialityDelta field explicitly to reflect that. The biggest single downward revision from March is at the 10-year horizon, where Cloud dropped from 4.62 to 2.95, and Big Pharma from 4.20 to 3.05. Not because the sectors got worse. Because the engine now knows more about how much of the future has already been paid for.

Second, Enterprise SaaS Legacy is now the single worst industry the engine covers, ranked 30th out of 30. It was already at -19.79% for H1. The engine says the multiple is still too high and the second half of 2026 through 2027 will keep grinding it lower. This is a specific, falsifiable claim. IGV was at approximately 82 at time of publishing.

Faster case, current trajectory — same 30 industries with AI moving faster

The Faster scenario asks what changes if AI capability improvement accelerates from today's pace. The pattern is amplification without re-ranking. The top of the matrix goes from 2.68 to 2.84 (Biotech-AI); Chip Design from 2.61 to 2.84; Cloud from 2.58 to 2.73. The bottom of the matrix softens rather than deepens. In Faster, Media at 5 years improves from 0.73 to 0.92, and Comm & Prof Services from 0.69 to 0.90. The engine says the negative sectors lose less in a faster-AI world because they participate more as consumers of AI capability, even as they get compressed as producers.

The bold bet the engine is currently making

Setting the caveats aside for a moment, here is what a person looking at the July 3 base case would take as the engine's actual view of the next five years.

AI drug discovery, AI-first cloud platforms, and the chip stack are the three industries most worth being long in 2026-2031. Not because the story is loud. Because the productivity gains show up in a place that is verifiable — new drug approvals per R&D dollar for biotech, capex-to-revenue efficiency for cloud, and unit economics that get cheaper every silicon generation for semis. The engine has these three industries clustered at the top in March, still clustered at the top in July, and will almost certainly still have them clustered at the top when the next quarterly refresh runs in October.

Enterprise SaaS Legacy is the single most short-worthy industry. The Salesforce, Oracle, Adobe, Intuit, ServiceNow archetype was built on the assumption that its per-seat pricing model was a moat. The engine says the AI-native competitors are eating the seat count from below, the platform assumption is eroding from above, and the multiple compression has three to five years of runway left. This is not a March call that got walked back; this is a March call the engine kept and is now more confident in.

Utilities and Communication & Professional Services stay at the bottom. The AI-datacenter narrative for utilities has always been the wrong end of the value chain — the money flows to behind-the-meter dedicated generation, not to publicly regulated rate-based utilities. Consulting and marketing services are the most directly language-work-substitutable sectors in the economy; the productivity dividend gets kept by the buyer, not the seller.

Insurance is the highest-conviction structural buy the engine has been early on. It missed at 5 months (KIE -0.83% in H1 2026). It might miss again at 12 months. But the engine put insurance at #4 in the base case at 5 years and it is the one call where the deployment evidence at Allianz, Baldwin, HUB, and AIG is running ahead of the regulatory friction and the tape. This is the wager the engine is putting the most calibrated conviction behind.

What the engine is most likely to be wrong about

An honest forecaster names the parts of the model that are least resilient. Four of them.

The consumer-versus-labor resistance split is one calibration cycle old. The v25 engine split what used to be a single "public resistance" field into two anti-correlated fields — one for consumer resistance to AI-delivered products and services, one for labor resistance to AI-driven displacement. The scores were set on one round of research and have not been backtested against actual union-vote data, actual class-action-adoption data, or actual consumer-refusal data. If either of those two fields is miscalibrated by more than half a point, the ranking of about six industries changes.

The capex-absorption-risk field has never faced a real hyperscaler earnings-miss quarter. The v26 engine added capexAbsorptionRisk to reflect the possibility that AI capital expenditure races ahead of AI revenue in a way that eventually compresses ROIC at the hyperscalers. The field currently has Cloud Platforms at 0.75 (high risk). If Q3 2026 hyperscaler earnings show that capex is being absorbed cleanly, the field is too pessimistic. If Q3 shows the first ROIC compression, it is roughly right. Either way, this is the single most refresh-sensitive number in the current matrix.

The Utilities call is the largest deviation from consensus in the whole matrix. A lot of thoughtful people think the AI datacenter buildout drives a utilities supercycle. The engine says otherwise, in the strongest possible terms — Utilities is dead last on the 5-year horizon at 0.70. If the engine is wrong about where the datacenter capex flows in the utility rate base, this is the call that will be pointed at first. It is currently the call the engine holds with the most confidence, which means it will also be the most instructive when it gets graded.

The engine is a five-year-and-beyond machine forced to also render 1-year and 2-year numbers. The short-horizon calls are the least reliable in the whole matrix. Anyone using the 1-year multiplier as an actual trading signal is asking the engine to do something it was not built to do. The direction of the 1-year call is usually right (72.7% hit rate on the March file). The magnitude is unreliable at the annual level and gets more reliable as the horizon lengthens.

Audit update — added July 3, 2026, the evening of publication

This page went up this morning with Utilities dead last at five years, and the section above calls it the engine's highest-confidence deviation from consensus. By this evening the call was dead, and I am the one who killed it.

The engine shipped a new attribution layer this afternoon. It decomposes every score into its components, and the first thing it showed me was that the Utilities ranking was not the regulatory-drag story I had been telling. It was driven almost entirely by cross-effect penalties: four pairs in the cross-effects matrix, calibrated in February, that had never been revisited while every other Utilities input got recalibrated five separate times. Three of them encoded a thesis the engine itself had already rejected, that more efficient chips would shrink power demand. The fourth coded hyperscaler self-generation as money leaving the utilities sector, when the marquee deals — Microsoft with Constellation, Amazon with Talen — are twenty-year contracts signed with utilities-sector companies. That is revenue moving around inside the sector, not out of it.

Correcting those four pairs moves Utilities from 27th to 18th of 30 at the five-year horizon, from 0.70 to 2.07. The new last-place holder is Communication & Professional Services, which was already in the bottom two and whose bear case does not depend on any February fossil.

The morning tables above stay exactly as published and will be graded exactly as published. That is what the frozen files are for. The live engine now carries the corrected matrix, and the October refresh will render from it. The full autopsy is in the calibration record: calibrations/v29.1-utilities-ce-audit.md and the v29.1 entry in calibrations/changelog.md.

A forecaster who cannot show you his mistake the same day he finds it should not be asking you to trust his hits.

How to grade this page in July 2027, July 2028, and beyond

Every number on this page is anchored to a preserved file. The March 4, 2026 numbers come from scenarios-computed_2026-03-04_v21-or-earlier_first-published-forecast-anchor.json. The July 3, 2026 numbers come from two exports: scenarios-export_2026-07-03_base-case_current-trajectory_public-post-anchor.csv and scenarios-export_2026-07-03_faster_current-trajectory_public-post-anchor.csv. All three files are frozen. They will not be edited. When July 2027 arrives, they will still say what they say today.

The grading protocol is simple. Pull the price of each named US sector ETF proxy on July 3, 2026 (or the March 4, 2026 close for the March forecasts). Compare to the same ETF's price one year, two years, three years, five years, and ten years later, dividend-adjusted. Compare that ratio to the multiplier the engine printed. The engine wins if the direction matches. It wins big if the magnitude does too. The results get published as they come in, on this page.

Published July 3, 2026. Next scheduled grade: October 3, 2026 (quarterly refresh). Underlying engine at the AI Market Cascade matrix.