딥 리서치가 내 투자 방식과 개발 방향을 바꿨다

These days, "deep research" feels less like a feature and more like its own product category. Services like OpenAI, Gemini, Perplexity, and Claude have moved well beyond simple Q&A — they now crawl the web, gather sources, and synthesize everything into a structured report. What matters now isn't who gives the longest answer, but who finds the best information, connects it most coherently, and verifies it most rigorously.

I believe this trend will only accelerate. As model capabilities improve, deep research improves with them — not just in writing quality, but in the ability to find relevant sources, filter out noise, and surface connections between disparate pieces of information. Personally, I think it's only a matter of time before deep research evolves from a convenience feature into something closer to a personalized, institutional-grade research service.

Using It Like a Research Assistant

Since deep research tools became widely available, I've been using them not as a simple search engine, but as a research assistant that helps me narrow down investment ideas. As a student, I couldn't invest large amounts — but I found the approach genuinely useful and saw satisfying results more than once. It helped me generate what I'd describe as "meaningful pocket money returns."

That said, the point here isn't to brag about profits. If anything, it's the opposite. Risk management matters far more. Deep research is a tool that improves your odds — it's not magic that eliminates losses.

So I don't use it as a "buy immediately because the AI recommended it" tool. I use it strictly to reduce the time I spend on pre-decision research and to surface evidence I might have otherwise missed. In practice, real investing requires stop-loss discipline, position sizing, avoiding overconcentration, checking counter-arguments, and revisiting recent developments. Skip those, and even excellent research becomes dangerous.

That said, I genuinely believe this kind of approach is one of the best starting points for a complete beginner who wants to get into investing. Rather than entering a position purely on gut feeling with no context, it gets you into the habit of reviewing multiple sources, articulating your reasoning, and examining the case against your thesis. It doesn't replace studying — but there's a meaningful gap between "investing with no framework" and "investing based on reasoned evidence."

My Workflow

My process is fairly straightforward. I give the deep research tool specific inputs: the current date, target return, acceptable risk level, investment horizon, preferred sector or market, and growth criteria.

For example, something like:

"Based on today's date, identify stocks with strong probability of returns within 1–3 months. Prioritize growth potential while keeping risk at medium or below."

I then generate a thorough report, and pass that report to a stronger reasoning model, asking it to compress the candidates down to 1–3 of the most compelling picks.

Here's the key insight: generating more candidates doesn't increase accuracy — it often dilutes it. In my experience, the most effective approach was constraining the output to 1–3 picks. Telling the model to select only from within the report (no injecting external knowledge), and requiring it to provide both the rationale and the counter-argument for each pick, sharpens the output considerably. When the report is long and document-heavy, giving the second-pass reasoning model enough tokens to work with also made a noticeable difference.

Cross-Comparison: Going One Step Further

Taking it a step further makes it even more powerful. Rather than relying on a single deep research output, cross-comparing results from two or three different models proved significantly more effective.

Each model has slightly different tendencies — what sources it gravitates toward, what evidence it weights most heavily, how it structures its conclusions. Relying on one model means accepting its blind spots and biases wholesale. Looking at multiple outputs side by side makes it much easier to identify the arguments that consistently appear across models versus the ones that diverge. This approach doesn't necessarily improve raw accuracy — it helps you better judge the strength and reliability of the underlying evidence.

In practice, this workflow produced real results. Stocks like NVIDIA, SanDisk, IonQ, SK Hynix, and SOXL were among the candidates generated through this process, and they resulted in actual gains at the time. But none of this should be taken as a success formula. What matters isn't which stocks appeared — it's that I entered those positions based on evidence. I reviewed risk scenarios and counter-arguments alongside each thesis, and that process is fundamentally different from "it looked good so I bought it."

Why It Works Especially Well in Finance

What I find particularly interesting is how much stronger this approach is in economics and finance specifically. These fields are inherently dependent on information quality, recency, and contextual interpretation — and deep research compresses all three quickly.

Rather than simply naming "stocks that look promising," it can show you why a conclusion was reached, what catalysts are driving it, what risks exist, and what sources informed the judgment. That transparency is what makes it useful rather than just convenient.

So I'm Building One Myself

That's why I'm now building my own deep research system. The goal is simple: not a service that wraps LLM APIs in a polished UI, but a genuine open-source deep research system that actually crawls the web, aggregates sources, cross-validates information, and produces trustworthy reports.

The benchmark I've set for myself is consistently clearing 60+ points on deep research benchmarks — performance approaching SOTA-level.

I've been almost exclusively focused on this goal for the past month, and I'm starting to see promising results. It's not public yet, but the direction is getting clearer. What I'm building isn't a "search chatbot." It's a general-purpose deep research system that genuinely retrieves, aggregates, and cross-validates to produce reports you can actually trust. Once the foundation is solid, I plan to follow up with finance-specific enhancements.

I'm convinced this space will become significantly larger. As models improve, deep research improves with them. It's already powerful — and it will only get more precise. That's why I don't think of deep research as just "AI that answers well." Designed right, it's a tool that changes how people investigate. Especially in fields like economics and finance — where information directly connects to money — that shift feels even more significant.

I see this not as a simple feature, but as one of the most powerful knowledge infrastructure tools an individual will have access to.

And I'm building it myself.

The bar is high. 60+ on benchmarks. Open-source deep research approaching SOTA. And eventually — an environment where anyone can access high-quality, personalized research.

That's the direction I'm moving. Every day. Once enough progress accumulates, I plan to release it publicly as open source.

※ This post reflects personal experience and observation only, and does not constitute investment advice. Any tickers mentioned are examples only. All investment decisions and their consequences are the sole responsibility of the individual investor.