Hands-On Test: WorkBuddy x DeepSeek V4.1 Flash for Stock Research — Don't Overlook These 4 Things

We ran a hands-on financial research test on WorkBuddy × DeepSeek V4.1 Flash. The focus wasn't on scoring it, but on seeing which parts of the research process it can save ordinary investors time on, and which checkpoints still must be handled by the investor themselves.

Here's the takeaway upfront: WorkBuddy can push data collection, draft calculations, and research leads forward very quickly. But whenever the answer involves actual judgment, time windows, attribution, units, and reasoning conditions still require manual verification. It's more like a research assistant that can rapidly lay out materials, rather than the final link that makes decisions for you.

Getting the Number Right Also Means Getting the Period Right

The most common source of misleading information in investment research isn't necessarily an outrageous number — it's a number that looks perfectly reasonable but covers the wrong time period.

In one market data case, the output placed chg_60d=-33.95% under the "year-to-date" performance column. In the same dataset, the actual year-to-date change was chg_ytd=+64.93%. Both numbers may have valid sources, but they answer completely different questions.

This is also where financial AI most easily lulls people into a false sense of security: when numbers, tables, and explanations appear together, it's easy to assume they are naturally consistent. In reality, as long as headers, body text, and data fields aren't aligned, any subsequent comparisons and conclusions lack a reliable foundation.

When you use AI to look at stock prices, returns, or growth rates, confirm three things first:

1. What time period does this number cover;

2. Whether "year-to-date," "past year," or "past quarter" in the text matches the original data field;

3. Whether the two sets of numbers being compared share the same starting point in time.

A number can be real, but placed in the wrong time frame, the conclusion will still be distorted.

Mentioning Something Doesn't Mean Attribution Is Complete

The second common issue is mistaking "appearing in the same paragraph" for "having a clear causal relationship."

In a case involving Tencent's earnings report footnotes, the footnotes covered multiple AI products or related items simultaneously. The output attempted to use this to explain a discrepancy of approximately RMB 10.5 billion, but the existing disclosures were insufficient to attribute that aggregated figure solely to any single item.

For investors, this distinction matters. A company's revenue, costs, investments, and impairments are often presented by segment, business line, or on an aggregated basis. A product name appearing in a footnote only indicates it may be related to that disclosure — it cannot automatically prove it is the sole cause of a particular change in amount.

So when you see AI making claims like "a certain business generated X in revenue" or "a certain product dragged down profits by Y," it's best to ask four follow-up questions:

• Which entity or segment does this amount correspond to;

• Is it revenue, cost, investment, or impairment;

• Does the disclosure state a confirmed fact, or is it management's explanation;

• Is there sufficiently granular breakdown to rule out other possible contributing factors.

When there is no line-item disclosure, the safer statement should be "this is a lead worth further verification," rather than writing the lead up as a conclusion.

A Single Conversion Can Be Off by an Order of Magnitude

Many people hand calculation tasks to AI because it seems the most objective. But the risk in financial calculations often lies not in the formula itself, but in details like units, exchange rate direction, and multipliers — things that can easily be glossed over in a single sentence.

In one exchange rate case, the output wrote up the foreign exchange calculation for ¥10 billion as approximately $100 million in losses. When recalculated using the conversion rate applied at the time, the result was approximately $10.2564 million. The issue isn't complex, yet it's enough to throw off any subsequent judgment by an order of magnitude.

The reminder for investors from this type of issue is straightforward: there's nothing wrong with having AI do calculations, but don't just look at the final result. Retain the original currency amount, exchange rate direction, unit multipliers, formulas, and time points, so you can tell whether it "produced a number" or "completed a verifiable calculation."

Especially when dealing with billions, millions, basis points, percentages, foreign currency conversions, and fiscal year conventions, a single recalculation is often worth more than a fluent paragraph of explanation.

Getting the Main Calculation Right Doesn't Mean the Research Is Done

In our testing, some questions had a main calculation that was correct. But when the answer went further to add supplementary ratios, cross-table bridging, period linkage, or extended a methodological example to a real company, new verification points emerged.

For example, the definition and sample calculation of working capital can be explained clearly, but that doesn't mean it already constitutes a complete working capital model for a real company. A real company also requires consideration of disclosed line items, accounting classifications, reporting periods, and non-recurring items. Similarly, even if the main formula in a valuation or financial calculation is correct, it doesn't mean the investment judgment derived from it has met sufficient conditions.

This is precisely where AI fits best in investment research: it can push the research to the "worth looking into further" stage, but it cannot declare on your behalf that "the research is complete."

When AI starts offering judgments like bullish, bearish, overvalued, undervalued, or "worth watching," at minimum the following should also be provided:

• The time horizon the judgment applies to

• The data and valuation methodology used

• What facts would invalidate the current conclusion

• What risk appetite this judgment suits, and who it doesn't suit.

Without these conditions, a conclusion is more like a research hypothesis than an actionable investment decision.

How Should Ordinary Investors Use WorkBuddy

If you position WorkBuddy correctly, it can still significantly lower the barrier to getting started with research.

Step one: use it to quickly lay out materials. Earnings highlights, industry background, historical data, draft calculations, and areas that need follow-up questions can all be organized by it first.

Step two: verify the parts that most affect your judgment. Prioritize checking time windows, units, entity scope, original sources, and intermediate calculations, rather than reviewing every sentence word by word.

Step three: rewrite conclusions as "conditional judgments." Instead of accepting "this company is worth buying," retain "if revenue growth, profit margins, and valuation assumptions still hold, this thesis is worth continuing to track."

Used this way, AI gives you not an answer that replaces your judgment, but a faster starting point for entering a research-ready state.

Conclusion: It Can Accelerate Research, but Shouldn't Skip It

The value WorkBuddy demonstrated in this test is connecting the tedious steps of collection, calculation, and explanation more quickly. For self-directed investors, that's already very useful.

But the most critical part of investment research usually isn't "can it write out an answer," but "can the numbers, attribution, and judgment conditions in that answer withstand scrutiny." Therefore, the more appropriate expectation is not to have AI make decisions for you, but to let it free up your time for the parts that truly require judgment.

nashnova: Making Research Go Beyond a Single Answer

WorkBuddy can get research started faster; the value of nashnova Agent is in making research go beyond a single answer. You can push a question forward as a Task on an ongoing basis, accumulating data, calculations, evidence, and conditions to verify in a Workspace, and continue updating when new earnings reports, new data, or counterexamples emerge.

For independent investors, what truly matters isn't getting more conclusions, but maintaining a research chain that is traceable, verifiable, and corrects itself as information changes. nashnova doesn't make investment decisions for you, but helps you actually keep that chain running.

Evaluation Notes

This hands-on test is based on nashnova's WorkBuddy × DeepSeek V4.1 Flash records from September 22, 2026. The sample consisted of 386 historical financial questions, of which 385 were completed. This article is for research and educational purposes only and does not constitute investment advice.