Presentation 2.0: Deep-Dive Equity Research with ChatGPT:
Creating Analyst-Grade Reports (and Beyond)
Three months back I published my presentation on the same topic here
I prepared the following presentation at the request of a group of friends who are interested in learning how to use ChatGPT. This time, I examined ChatGPT through first-principles reasoning, a process that has markedly enhanced my understanding and will enable me to use the tool more effectively going forward.
You can download the entire presentation from here
For the past decade, I have been fortunate to receive both direct and indirect mentorship from three well-wishers, from whom I have learned a great deal. Their guidance was instrumental in shaping this presentation, distilled into three core principles:
• Always think from first principles. [VT]
• The person delivering the presentation learns the most. [DS]
• Be curious. [SB]
1. ChatGPT under the bonnet: nothing but next‑token maths
At its mathematical core GPT‑style models (o3, o3 Pro or otherwise) do one thing only—predict the most probable next token [word]. The “magic” is just probability piles stacked 100 billion‑high. Because it has no persistent facts or opinions—only statistical echoes—whatever you feed the model in your prompt acts as the gravity that bends its token stream. The diagram on page 6 makes this point visually: the “GPT engine” shoots tokens down a conveyor and the Prompt Dial narrows or widens the beam.
Why this matters: if you pre‑load sloppy or missing instructions [prompt], the model will confidently predict nonsense; supply tight instructions and you’ll get focused, defensible analysis.
2. What you can control (and should)
The presentation’s score‑card on page 9 lists every lever that shapes output quality.
Condensed, the user‑controllable levers are:
Everything else (system rules, developer rules, the hidden self‑planning layer) is off‑limits, so double‑down on the seven knobs above. See the control‑stack graphic on page 13.
3. Prompt vs. Research Plan
Think of a Prompt as all the instructions you feed into CHATGPT. The clearer the prompt, the better the result.
A Research Plan is the full road‑trip itinerary: fuel stops, detours, hotel bookings. Skip it, and ChatGPT guesses the route [which may be completely different from what you wish to take]. The side‑by‑side car analogy on page 10 underlines this difference.
4. The Five Elements of a Complete Prompt
The five icons repeated across pages 19‑24 give you a ready cookbook:
Fail to include any one of these and, as the caution table on page 25 shows, you risk logic gaps, fluffy narratives, or hallucinated data.
5. The Four‑Step Method to Generate a Research Plan
A single slide (pages 17 & 54) distils the workflow. Here it is in plain language:
Draft a preliminary prompt
Include the five elements above.Ask ChatGPT to “create a better prompt”
Leverage the model’s own self‑critique to tighten scope and clarity.Request: “Prepare a research plan which I can submit to CHATGPT o3 [not human] to achieve the purpose in the prompt.”
The o3/o3 Pro reasoning model breaks the problem into research modules and lists sources.Run Deep Research
Upload or paste the plan and simply say: “Execute the research plan.”
Iterate: reject any weak sections, revise the plan, and rerun the process. Ensure that you thoroughly review the prompt and the research plan generated by ChatGPT, making all necessary adjustments. Because ChatGPT’s deep-research function produces reports of 20–40 pages, allocate sufficient time to refining both the prompt and the research plan; otherwise, you risk spending time on a lengthy report that fails to meet your objectives.
The funnel graphic on page 59 illustrates how each tweak narrows uncertainty.
6. Extra power‑ups (optional but valuable)
High‑impact directive phrases – “Be critical”, “Think from first principles”, “Quantify” (see page 46) add analytical guard‑rails.
Hallucination guard‑rails – force tables, require citations, cross‑check maths (checklist on page 64).
Multiple personas – run the same question through CFO, Short‑Seller, and Industry Expert personas to stress‑test conclusions (page 42).
Key‑Takeaway Table
But does the research worth relying on? I mean the prompt we put is what matters but to add the data source and other important material. We itself need to research and use our thinking.