What 4,523 AI investment forecasts revealed about model disagreement

Russlan

Jul 23, 2026

One market. Multiple analytical lenses. One inspectable consensus.

By Russlan Ramdowar, Founder of iPulse AI

The hardest part of investing is rarely clicking Buy. It is deciding what deserves your capital—and staying rational once your own money is involved.

Investing has never been more accessible. A person with a phone can buy a fractional share of a global company in seconds.

Good judgment has not become equally accessible.

The moment money becomes personal, analysis stops being purely analytical. We anchor to the price we paid. We search for evidence that confirms what we already own. We sell because a red number frightens us, buy because a green chart excites us, and confuse activity with control.

The cost is measurable.

Morningstar’s 2025 Mind the Gap study estimated that the average dollar invested in U.S. mutual funds and ETFs earned 7.0% a year during the ten years ended December 2024. The funds themselves returned 8.2%. The 1.2-percentage-point annual gap was associated with the timing and size of investors’ purchases and sales.

That sounds small. Compounding says otherwise. If those two rates hypothetically persisted for 30 years, $10,000 would grow to roughly $76,000 at 7.0% and $106,000 at 8.2%—a difference of about $30,000 created without changing the underlying funds. That illustration is not a forecast. It is a reminder that behavior becomes arithmetic.

Professionals are not immune. In the S&P Dow Jones Indices SPIVA U.S. Year-End 2025 scorecard, 79% of active U.S. large-cap funds underperformed the S&P 500 in 2025. Over five years, the underperformance rate was 89%.

Yet disciplined analysis can also compound spectacularly. Berkshire Hathaway’s official 2025 annual report records a 19.7% compounded annual gain from 1965 through 2025, compared with 10.5% for the S&P 500 with dividends included.

That history is not a recipe, and it does not mean anyone can reproduce Warren Buffett’s results. It demonstrates something more durable: a consistent analytical framework, emotional restraint, and long-horizon compounding can produce a radically different outcome.

Capital allocation is how savers meet builders

The world contains people and institutions that hold capital, and people and companies that can turn capital into products, infrastructure, medicines, software, jobs, and new productive capacity.

Capital allocation connects them.

Public markets are one of the most accessible versions of that bridge. The SEC’s latest rolling statistics count 3,600 U.S.-domiciled exchange-listed companies and another 1,139 foreign-domiciled listed companies in its reporting population.

Behind every ticker is a real capital-allocation question:

Which businesses deserve more resources?

Which industries are building something the world will need?

Which management teams can convert capital into durable cash flows?

Which prices already assume a future too optimistic to survive?

Which risks are visible, and which are being ignored?

Choosing well can grow personal wealth while directing capital toward productive companies. Choosing badly can destroy capital, reward weak economics, or leave money trapped in a persuasive story.

The opportunity is enormous. So is the research burden.

No individual can continuously read every filing, earnings call, macroeconomic release, price history, valuation signal, competitive threat, and geopolitical event across thousands of securities—then apply the same standard to every asset while remaining emotionally detached.

This is where AI becomes genuinely interesting.

The goal should not be to surrender responsibility to an opaque machine. That would replace one emotional shortcut with one technological shortcut.

The real prize is to outsource the repetitive, cold-blooded analytical work while keeping the evidence, assumptions, disagreement, and final responsibility visible to the human investor.

A machine should not eliminate judgment. It should make disciplined judgment easier to practice.

What our latest market-wide workflow discovered

In July 2026, iPulse AI completed its latest July Market-Wide Deep Analysis Workflow: a five-year, multi-asset research run spanning equities, cryptocurrencies, commodities, indices, and currency pairs.

The workflow produced 4,523 individual model-asset ratings across 377 assets.

In 290 of those assets—76.9%—at least one AI voice was positive and at least one was negative.

Then we found something even more revealing.

Of the 179 assets that finished with a NEUTRAL consensus signal, 174 contained both positive and negative ratings.

That is 97.2%.

Neutral was usually not an absence of opinion. It was a collision of opinions.

A final rating can look calm while the system underneath is shouting.

This matters because the investment industry keeps asking AI for a single answer: buy, hold, or sell. A single answer is easy to display. It is also capable of deleting the argument that produced it.

Our data suggests a better question:

What if disagreement is not noise around the signal—but part of the signal itself?

The four numbers worth remembering

377 assets in the analyzed research snapshot

4,523 individual model-asset ratings

76.9% of assets contained opposing positive and negative views

97.2% of neutral consensus signals contained both sides

Figure 1. A model-rating distribution was classified as “opposing” when it contained at least one STRONG BUY or BUY and at least one PARTIALLY SELL or SELL ALL. Source: iPulse AI Data Platform; July 2026 five-year original research snapshot.

What we actually analyzed

We used the latest active record for each asset in iPulse AI’s original, non-adjusted, five-year scoring snapshot recomputed in July 2026.

The snapshot contained:

348 equities

14 cryptoassets

5 commodities

5 indices

5 currency pairs

Of the 377 assets, 376 had ratings from 12 AI voices and one had 11, producing 4,523 ratings in total.

We grouped STRONG BUY and BUY as positive ratings. We grouped PARTIALLY SELL and SELL ALL as negative ratings. An asset entered the “opposing views” group only when both appeared in its rating distribution.

These are model forecasts, not realized returns. The analysis tells us how the system’s views were distributed—not which forecast will ultimately be correct.

That distinction is essential.

Why ask several AI advisors instead of one?

A single AI can produce a polished answer. That does not mean it has examined the problem from every investment perspective.

A value investor asks whether the price is supported by durable economics. A macro strategist asks how the asset behaves under changing inflation, rates, liquidity, and growth. A contrarian searches for what the crowd may be missing. A technology-focused thinker asks how innovation changes the size and structure of the opportunity. A strategic-power lens examines regulation, incentives, competitive control, and institutional influence.

iPulse AI turns those different disciplines into governed advisor configurations.

Recognizable persona families—such as Warren Buffett, Ray Dalio, Michael Burry, Elon Musk, and Machiavelli—label different algorithmic analytical frameworks. They are not endorsements, replicas, or claims that those individuals produced the analysis. They are a practical way to make each framework’s intellectual starting point understandable.

The wider platform supports more than 100 analysis frameworks. For the snapshot in this article, up to 12 scored AI voices examined each asset. The workflow then:

applies different analytical lenses to the same security;

produces independent reports, scenarios, risks, and forecast paths;

normalizes the forecasts so assets can be compared on a common foundation;

measures agreement, direction consistency, and forecast dispersion;

incorporates return magnitude, dividends, volatility, and risk pressure;

creates one ranked consensus signal without deleting the underlying reports.

This architecture matters because an average is useful only when you can still inspect what was averaged.

“Neutral” can hide a knife fight

Consider three familiar names in the snapshot:

NVIDIA: 6 positive, 2 neutral, and 4 negative ratings; final signal: NEUTRAL

Amazon: 8 positive, 1 neutral, and 3 negative ratings; final signal: NEUTRAL

Tesla: 4 positive and 8 negative ratings; final signal: NEUTRAL

Now consider the most polarized case in the batch:

Quantum Computing Inc. split exactly down the middle—6 positive ratings and 6 SELL ALL ratings. Its final signal was NEUTRAL.

Calling that “no opinion” would be absurd. There was plenty of opinion. There was no agreement.

Figure 2. Selected examples illustrate different internal paths to a final consensus label. Rating counts are model outputs, not investment recommendations. Source: iPulse AI Data Platform.

The examples also show why a simple vote count is not the complete system. The final consensus score incorporates forecast magnitude, directional consistency, risk information, and other configured scoring inputs. Eight positive ratings do not automatically create a BUY, just as one negative rating does not automatically cancel eleven positive ones.

The label is a summary.

The distribution is the evidence behind the summary.

Even a bullish verdict often contained a bear

The disagreement was not confined to neutral signals.

The batch contained 174 assets with a final positive signal—BUY or STRONG BUY. 101 of them, or 58.0%, still contained at least one negative rating.

The reverse was also true. Of the 24 assets with a final negative signal, 15, or 62.5%, still contained at least one positive rating.

Compare three positive signals:

Bitcoin: 12 of 12 ratings were positive; final signal: BUY

Ethereum: 10 positive and 2 negative ratings; final signal: BUY

Intuit: 11 STRONG BUY ratings and 1 SELL ALL rating; final signal: STRONG BUY

The headline label alone makes these cases look more similar than they are.

Bitcoin’s positive signal emerged from unanimity. Ethereum’s emerged with a minority objection. Intuit’s emerged with one extreme dissenter.

For a serious researcher, those are three different information states.

Disagreement changed the shape of the forecast

We then compared the 290 assets with opposing views against the other 87 assets in the snapshot.

The opposing-view group had:

More than twice the average annual-return dispersion: 7.18 percentage points versus 3.46

Lower average direction consistency: 0.731 versus 0.852

Higher average risk pressure: 52.2 versus 42.6

A much higher neutral-signal share: 60.0% versus 5.7%

A smaller average absolute consensus score: 108.4 versus 206.3

Annual-return dispersion here is the median absolute deviation of the AI voices’ projected annualized returns. In plain English: when the voices disagreed on direction, their estimates also tended to spread much farther apart in magnitude.

Figure 3. Descriptive comparison within the July research snapshot. The groups were not randomly assigned, so these differences should not be read as causal effects.

That does not prove disagreement causes risk. It does not prove a dissenter will be right. It says something more modest and more useful:

The presence of opposing views identified a materially different forecast profile.

If a platform exposes only the final signal, that difference becomes hard to see.

What this gives someone making a decision today

Imagine that you are considering NVIDIA, Bitcoin, a pharmaceutical company, or a defensive dividend stock today.

The normal workflow is fragmented. You search the web, open several charts, read a few articles, watch a video, ask a general-purpose chatbot, and slowly assemble a thesis from sources that use different assumptions and time horizons.

iPulse AI is designed to turn that process into one repeatable research sequence:

Discover: begin with Top Picks, where assets are ranked un

[truncated for AI cost control]