I Tested 14,400 AI Forecasts at Two Horizons. Here Is the Hidden Pattern I Found.

Originally published in DataDrivenInvestor on August 23, 2026. Republished on iPulse AI on September 19, 2026. Figures and findings retain the original research dates; this is a historical research edition.

For years, one kind of email kept irritating me.

IBKR would send an alert: an analyst had changed a stock rating.

Upgrade. Downgrade. And little else.

More irritating still, the email seemed to arrive only after the market had already made a violent move.

My frustration was not with IBKR. The broker was relaying research produced elsewhere (usually a Reuters aggregator). It was with the strange theater around the rating itself.

An older JPMorgan rating may well have been reevaluated after new evidence arrived and intentionally left unchanged. A standing BUY can simply mean: we looked again, and it is still a buy. The screen did not make that process transparent, but it would be unfair to assume no review occurred.

That table was still obscuring two different variables.

Forecast vintage , or the as-of date, tells us when the analysis was produced and what evidence it could have known.

Forecast horizon tells us how far into the future its conclusion is meant to describe.

Vintage matters because it establishes what evidence was available. Yet horizon is the more fundamental ambiguity. Even two ratings produced from the same evidence can point in different directions if one evaluates the next twelve months and the other the next five years.

When someone writes BUY, until when? After a 20% gain, by next quarter, or over a long-term holding period? When someone writes SELL, are they rejecting the company’s long-term economics or warning that the next six months may be brutal?

A rating without a horizon is like a GPS direction without a marking on the final destination. It sounds decisive. It may even be technically correct. It just cannot tell you where you are going.

That irritation, along with similar inconsistencies in investment research tools, sent me on a mission to build my own research institute. What followed felt like a philosophical journey.

Fast-forward three years of continuous building, testing and iteration. Our team had generated and examined 14,400 AI investment forecasts across two horizons: one year and five years.

The experiment isolated the horizon. Every pair used the same forecast vintage and identical contributing forecast set.

And here comes the shocking part. Half the resulting ratings changed direction…

Not the model. Not the evidence. Not the asset.

The forecast horizon.

14,400 AI analyst opinions is almost impossible to picture

Imagine that every investment thesis takes only five minutes to read.

Reading 14,400 of these would consume 1,200 hours. At eight hours a day and five days a week, that is nearly seven months of reading, with no holiday, no sick day and apparently no meetings.

Yet the size of the archive was not the revelation. The horizon was.

We generated four research batches between April and July across 377 assets, with 10 to 12 separately configured investment frameworks examining each asset. Every framework produced a five-year path, evaluated at Year 1 and Year 5.

Those 14,400 contributing forecasts became 1,247 consensus ratings per asset per batch. The research universe expanded across batches, reaching 377 assets in the final batch.

Each consensus verdict must land in one of five ordered ratings:

• SELL ALL
• PARTIALLY SELL
• NEUTRAL
• BUY
• STRONG BUY

First, we asked the literal question: did the exact rating change when the horizon moved from one year to five years?

It did in 705 of 1,247 comparisons, or 57% .

Then we made the test harder. We collapsed the five labels into only three broad directions: negative, neutral and positive. A move from SELL ALL to PARTIALLY SELL would no longer count because both are still negative. A move from BUY to STRONG BUY would no longer count because both are still positive.

Even under that stricter test, 623 verdicts changed direction.

Half, rounded.

My first reaction was not excitement. It was suspicion.

Did we break a join? Swap the horizons? Let different advisor tasks enter the two verdicts? Three years of building an investment-research system teaches a useful reflex: a surprising result is guilty until proven innocent.

We checked the 1,247 pair keys, the contributing task IDs, the scoring outputs and the null rates. Every one-year verdict was paired with the five-year verdict built from the identical underlying forecast set. There were no duplicate pairs, no missing signals and no mismatched task sets.

The result survived the interrogation.

The rating did not change its mind. We changed the question.

Look closely at the transitions and something even more interesting appears.

The system did not stampede from panic to euphoria. Only 13 verdicts jumped from a negative one-year direction to a positive five-year direction.

The larger movements were quieter:

• 237 shifted from PARTIALLY SELL to NEUTRAL.
• 282 shifted from NEUTRAL to BUY.
• 73 shifted from SELL ALL to NEUTRAL.
• 323 negative one-year verdicts became either neutral or positive over five years.

This was not optimism sprayed over a longer chart. Temporary friction lost power while compounding mechanisms gained it.

Both evaluations used annualized average return. Comparing a one-year return with an unannualized five-year cumulative return would mechanically favor the longer period; that was not the comparison here.

Why should the horizon do that?

Because the horizon decides which facts are allowed to matter.

Over one year, an earnings miss can dominate. So can a budget freeze, a product delay, a rate shock, a regulatory decision or forced selling. A strategically healthy business can spend twelve miserable months trapped beneath a bad comparison or an expensive valuation.

Over five years, different forces enter the room: market-share gains, capital allocation, infrastructure build-out, falling unit costs, demographic demand, network effects or a business model that becomes more valuable as adoption spreads.

The pattern made me reconsider how strongly a chosen horizon shapes a thesis. It does not establish that holding for longer produces superior realized returns.

Investment research keeps printing both answers with the same five words.

That is the hidden problem. BUY without an explicit horizon is not a complete opinion. It is the final line of an argument after the scale has been removed.

The effect returned in every batch

• Batch 3: 131 of 246 verdicts changed broad direction, or 53%.
• Batch 4: 152 of 312 changed, or 49%.
• Batch 5: 176 of 312 changed, or 56%.
• Batch 6: 164 of 377 changed, or 44%.

The percentage moved. The phenomenon did not disappear.

Across four expanding research runs, roughly four to six of every ten asset verdicts changed direction when the horizon changed.

Three assets, three ways the horizon changed the thesis

To see what the pattern meant inside recognizable investments, I separately inspected a later August recomputation of Batch 6. That snapshot is excluded from the headline denominator.

It contained nine cases that were negative over one year and positive over five. They did not tell one cheerful sector story. They showed three ways a longer horizon can alter the meaning of the same evidence.

Accenture: disruption now, integration later

Generative AI can automate coding, testing and analysis that consulting firms once sold by the hour. Corporate caution and public-sector pressure can delay projects before reinvention revenue arrives.

Our one-year model view reflected that friction: PARTIALLY SELL, with an annualized mean return forecast near -5%.

Then the horizon moved from one year to five.

Large enterprises do not install probabilistic systems as if they were office software. They have old data, incompatible systems, compliance duties and executives who need someone accountable when automation fails. AI may compress parts of Accenture’s labor model while expanding demand for integration, governance and organizational change.

The official numbers reveal the tension. Accenture reported 104 quarterly client bookings worth at least $100 million year to date, up 13%, and said it was seeing more large-scale AI transformation programs. It also reported 3% local-currency revenue growth and a 17% operating margin. Five days later, the company expanded planned fiscal-2026 repurchases to $7.5 billion, 62% above the prior year. Accenture Q3 FY2026 results , share-repurchase announcement

The five-year model view moved to BUY, near 12% annualized. The gap between -5% and 12% was not a mood swing. It was a race between two mechanisms operating at different speeds.

TotalEnergies: the oil cycle now, the power system later

Over one year, commodity prices, refining margins, regulation and geopolitics can dominate TotalEnergies. The model’s one-year signal was PARTIALLY SELL, near -2% annualized.

Over five years, the company starts to resemble more than a wager on the next barrel of oil. It becomes a portfolio of molecules, electrons, trading infrastructure and capital returns. Its ability to pair gas, renewables and flexible power may matter more as data centers demand electricity that is both abundant and dependable.

That mechanism is already visible. TotalEnergies has described AI as an accelerator of data-center power demand. A clean-firm-power agreement announced for France covers 1.5 terawatt-hours over 15 years, while a separate Data4 contract beginning in 2026 covers 610 gigawatt-hours over ten years. TotalEnergies 2026 climate presentation , Data4 power agreement

The five-year model moved to BUY, near 11% annualized. A map of competing horizons: commodity pressure now; infrastructure demand, portfolio transition and shareholder yield later.

Ethereum: opportunity cost now, network evolution later

Ethereum produced the most dramatic and least comfortable split.

The one-year model view was PARTIALLY SELL, near -5% annualized. The frictions included competition from faster chains, regulatory uncertainty, the opportunity cost of safer yields and the possibility that cheaper Layer-2 activity weakens fee burn on the base layer.

Over five years, the same scaling architecture can mean something else. Lower fees may weaken one source of scarcity while making the network more useful. The official roadmap still targets substantial increases in execution and data capacity, with Glamsterdam planned for the second half of 2026 and longer-term work focused on scaling, usability and hardening the base layer. Ethereum’s documentation is careful to say priorities and timing can change. Ethereum roadmap , 2026 protocol priorities

The five-year model moved to BUY, near 22% annualized.

Tempting? Yes.

Comfortable? Not remotely. The same result carried risk pressure of 93 out of 100. The long horizon enlarged the opportunity; it did not make the path safe.

Patience is not the absence of risk. Sometimes it is simply an agreement to remain exposed to risk for longer.

Long term is where weak assumptions go to hide

There is an obvious objection: stretch the horizon far enough and almost every difficult story can be made to sound hopeful.

I agree.

The five-year answer is not the wise answer floating above a noisy one-year market. It is a conditional answer with a longer chain of ways to fail. Management changes. Technologies disappoint. Regulation moves. Capital gets diluted. A company can run out of money before its supposedly inevitable future arrives.

And, as the old market warning goes, “the market can stay irrational longer than you can stay solvent.”

The line matters here because a forecast can ultimately be right and still be unusable. If your capital, mandate or psychology cannot survive the path, being early is simply another form of being wrong.

That is why every long-term thesis needs an invalidation condition.

For Accenture, the case weakens if autonomous tools bypass the integration layer faster than enterprises create demand for governance and change. For TotalEnergies, it weakens if transition capital earns structurally poor returns or regulation captures the upside. For Ethereum, it weakens if scaling activity fails to return economic value to the base asset or protocol and regulatory risk break institutional trust.

A forecast should not only say where an asset might go. It should say what must remain true long enough for it to get there .

The finding forced us to become more rigorous

This was the philosophical part for me.

I had started with a familiar annoyance: stale and fresh analyst ratings sitting beside one another as if BUY and SELL were timeless facts. Then our own system demonstrated just how much information a rating can erase.

The answer was not to abandon ratings. Compression is useful when hundreds of assets and thousands of forecasts must become comparable. The answer was to stop allowing the compressed label to travel alone.

Our research discipline now treats five fields as inseparable:

• Forecast vintage and market state: When was the analysis produced, what was publicly knowable and what price anchored it?
• Horizon: When is the claim supposed to mature?
• Near-term friction: What could dominate before the thesis has time to work?
• Compounding mechanism: What process is expected to create value over time?
• Invalidation condition: What evidence would tell us the thesis has broken?

That structure makes research less theatrical and more accountable. It also makes old forecasts useful. Instead of recording only whether a rating was right or wrong, a team can examine what the model knew, what it missed, which mechanism failed and whether the same reasoning error is recurring.

The first forecast supports a decision. The preserved reasoning can improve the next thousand.

The question hiding underneath every rating

When a rating alert appears now, my first question is no longer simply who upgraded or downgraded the asset.

I ask when the analysis was produced. What happened after it? Which horizon governs the conclusion? Can the portfolio survive the journey? What would prove the analyst wrong?

I began by trying to build a better rating system. I ended up learning that the rating itself is often too small for the truth it claims to contain.

Markets contain processes that mature over radically different horizons. Earnings revisions, policy decisions, product cycles, infrastructure build-outs and demographic changes all move at different speeds. Forcing them into one timeless column creates the illusion of contradiction.

So before asking whether an asset is a buy, ask the question hiding underneath:

By when?

The horizon is not metadata beside the thesis.

It is part of the thesis.

Method note: The four-batch figures are derived from original non-adjusted iPulse AI consensus-scoring records for Batches 3–6, dated April 11 through July 5, 2026. The analysis counted 14,400 unique contributing forecast task IDs and paired 1,247 asset-by-batch consensus rows at one-year and five-year horizons. Every pair used an identical contributing task set; there were no duplicate pair keys or null signals. The three named case studies use a separate Batch 6 recomputation dated August 20, 2026 and are excluded from the headline denominator. Percentages are rounded to the nearest integer. Model outputs are uncertain, are not realized performance and do not constitute investment advice.