Kimi Financial AI Tested: If the Valuation Math Works Out, Does That Make the Investment Case?

We ran a hands-on test across four common tasks: building a DCF, comparing technical indicators, explaining changes in operating profit, and reading an earnings report conclusion. The takeaway is straightforward: Kimi can push the first step of research forward very quickly, but it cannot complete the final step of judgment for the investor.

It excels at breaking problems apart, placing numbers into a structure, and surfacing contradictions you would otherwise have to find yourself. What still requires a human is verifying data timestamps, judging whether assumptions are reasonable, and distinguishing "what the model derives" from "what is actually true."

Core Conclusion: Treat Kimi as a Research Assistant, Not a Conclusion Machine

The easiest thing to overlook in a fluent AI analysis isn't the formula — it's the premise behind the formula. It can tell you how high a company's valuation appears, how an indicator has performed historically, or how much profit a particular business line may have contributed. But none of these are automatic answers for "you should buy" or "operations have already improved."

In this test, Kimi demonstrated practical value in organizing frameworks, completing calculations, and rapidly generating a first draft of research material. The role it is best suited for is helping you start research faster: producing an initial model, listing key variables, and flagging where further questions are needed. The final investment thesis still has to come back to original disclosures, data definitions, and your own judgment.

Case 1: The DCF Math Works, but the Real Disagreement Is Still in the Assumptions

We asked Kimi to produce a DCF valuation for Haitian Flavouring. The model output an intrinsic value of RMB 41.81 per share, compared to a market price of RMB 33.83. Looking at these two numbers alone, it's easy to form a first impression that "the current price looks attractive."

But scroll further down the spreadsheet and the issues emerge: roughly 69% of the enterprise value in this valuation comes from the terminal value. In other words, the model is highly sensitive to long-term growth, margin, and discount rate assumptions — no matter how precise the near-term forecasts are, they cannot mask the dominance of long-term assumptions over the result.

The sensitivity analysis further illustrates this point. When the discount rate is adjusted to 8.0%, the valuation is approximately RMB 34.16; when raised to 8.5%, it drops to approximately RMB 31.71. The apparent margin of safety quickly narrows or even disappears.

This is precisely where Kimi is most useful — and most dangerous — in investment research: it can assemble a valuation table very quickly, but it won't judge for you whether the long-term assumptions hold up. We also found that the output text had misread the discount rate range corresponding to a certain price. Getting the spreadsheet math right doesn't mean the interpretation of the spreadsheet is also correct.

For ordinary investors, the best use of this kind of output is not to memorize a target price, but to follow up with three questions: Where do the growth assumptions come from? Is the terminal value share too high? If the discount rate shifts slightly, does the conclusion still hold?

Case 2: Technical Indicators Have Answers, but That Doesn't Mean They've Given You a Winner

Technical analysis is well-suited for AI involvement because it can quickly calculate, organize historical results, and place different rules into the same table for comparison. But the real mistakes tend to come from forcing different samples into a single conclusion.

A user asked whether the 60-day Simple Moving Average (SMA) or Exponential Moving Average (EMA) is more valuable. The Agent wrote a backtest and produced returns, drawdowns, trade counts, and win rates. In the Shanghai Composite sample, SMA returned 22% and EMA 19%; in the S&P sample, SMA returned 24% and EMA 25%. The two sets of results point in different directions.

This means you cannot simply say "EMA is better" or "EMA is worse." Different markets, different time periods, and different sample weightings can all change the conclusion. AI can help you spot these differences faster, but if it only delivers one neat summary judgment, it actually obscures the information most worth investigating.

When using Kimi to compare indicators, consider asking it to preserve results broken down by market and by period, and to clearly state the sample definitions. First see where the differences come from, then decide whether to accept a single aggregate answer.

Case 3: The Model Tells You Where to Dig Deeper — It Hasn't Explained Actual Performance

Operational attribution is another type of task easily misled by a "sense of numbers." In one per-vehicle profit estimation, the model produced a profit increment of RMB 4,189.77 per vehicle, of which the overseas portion contributed RMB 3,997.12, or approximately 95.4%.

This is not a sentence you can drop directly into a company conclusion. It only shows that, given this set of inputs and assumptions, overseas-related variables dominated the model's output.

In reality, profit improvement may also be affected by product mix, raw materials, exchange rates, channel expenses, and accounting treatments, among many other factors. Kimi can quickly flag the signal that "the overseas assumption is critical," helping you determine what to check next; but it cannot use a single estimation table to prove that "95.4% of profit improvement came from overseas operations."

A contribution within a model is, first and foremost, an explanation of the model — not a definitive statement about the real world.

Case 4: Even a Single Earnings Conclusion Requires Unpacking the Profit Definition

Earnings report interpretation may seem simple, but it truly tests patience with definitions. In the case of POSCO's Argentine lithium business, operating profit was KRW 11 billion, while net profit was KRW -31 billion.

These two numbers are not contradictory. A positive operating profit suggests that the core operating segment may have improved; a still-negative net profit means that taxes, interest, foreign exchange, or other non-operating items may still be affecting the bottom line.

If AI compresses this information into "the company has turned profitable," it reads smoothly but has already jumped across profit levels. Similarly, words like "largest" and "leading" need to be matched with clear geographic scope and comparison benchmarks.

Kimi can help you read through earnings reports faster, but it cannot skip this step for you: when you see a strong conclusion, go back to the source to confirm whether it refers to operating profit, net profit, or some narrower operating metric.

How Can Ordinary Investors Use Kimi for Research?

Using Kimi in the "build the framework first, verify later" stage is usually more valuable than asking it to deliver investment conclusions directly.

First, have it build the model, then list the key assumptions. For example, when running a DCF, ask it to simultaneously show the impact of growth, margins, discount rate, and terminal value on the result — rather than just giving a single valuation number.

Second, have it rewrite a judgment into a verification checklist. When you see "overseas business drove profit improvement," follow up by asking: what segment data, what quarterly trends, and what company disclosures are needed to support this statement?

Third, have it preserve counterexamples. When comparing technical indicators or industry data, ask it to list samples that point in the opposite direction separately, rather than forcing everything into a single winner.

Fourth, have it state what would overturn the conclusion. The AI output most worth continuing to research isn't the one with the most confident tone — it's the one that can tell you "if this data point appears next, I would need to revise my view."

AI's Greatest Value Is Surfacing Questions Earlier

The value of financial AI tools like Kimi lies not in making conclusions for you, but in putting the questions worth verifying in front of you faster.

It can help you go from a blank page to a first-draft research framework: get the model running, align the numbers, and surface contradictions. The work that follows still belongs to the researcher: checking sources, examining assumptions, finding counterexamples, and revising views as information changes.

Being able to calculate a valuation is where research begins; being able to explain assumptions and withstand counter-evidence is what brings it closer to a complete investment study.

nashnova: Turning a One-Time Answer into Sustainable Research

The role of the nashnova Agent is precisely to turn this kind of caution from a one-time reminder into an ongoing research process: saving assumptions, materials, and open verification questions around a Task; feeding new data into the evidence base, then revisiting whether the original conclusion still holds.

For independent investors, nashnova doesn't decide your trades for you; it helps you place the model's answers, original sources, falsification conditions, and subsequent updates into a single research chain.

Testing Methodology

This article is based on a hands-on test completed by nashnova on September 22, 2026. The test subjects were Kimi K3, Kimi Hosted Agents, and the financial suite, with a sample of 500 questions. In this round, all 500 questions were completed; out of 37,890 checkpoints, 1,580 were not adequately supported, corresponding to an atomic-level evidence gap rate of 4.17%, with 310 questions showing at least one composite evidence gap.

These figures describe the evidence boundaries within the scope of this particular test. They do not equate to Kimi's overall error rate, a public product capability ranking, or any investment conclusion. The DCF, technical indicators, operational attribution, and earnings data in this article's cases all come from the same test record. No judgment has been made on Kimi's full range of public capabilities, data sources, security, stability, or efficiency.

nashnova provides market intelligence and research support. It does not provide personalized financial advice. This article is for research and educational purposes only and does not constitute investment advice.