Best Time Frame 1mo vs 3 mo?
Prompt:
In Indonesian market, for Gemini Deep Research (no other API/tools), will it be easier to predict 1mo stock movement or 3mo? Why? By easier, I mean more accurate from the side of direction conviction.
Answered by Gemini-GPT-Claude (GGC) on 2026-05-08:
Gemini 3.1 Pro: For Gemini Deep Research, the 3-month horizon offers significantly higher directional conviction and accuracy in the Indonesian market compared to a 1-month outlook. While the “easier” path in trading is often subjective, the way an LLM-based research tool processes information favors the medium-term fundamental “signal” over the short-term market “noise.”
GPT 5.5 Pro: If “easier” means higher directional conviction — up / down / sideways — using Gemini Deep Research only, then my answer is:
3-month movement is usually easier to predict with higher conviction than 1-month movement in the Indonesian market.
Claude Opus 4.7: 3-month wins on directional conviction because Gemini Deep Research’s information base (analyst consensus, macro data, earnings outlook, commodity trends, fund flows) is structurally aligned with quarterly horizons. You simply get more independent signals pointing the same way, which is the definition of higher conviction.
LLM Council Method
Step 1A: Open Router Fusion (LLM Council)
Ask this to Open Router Fusion, let the multiple LLMs conclude for us automatically:
Objectives: I want to create a 15 stocks portfolio that is SAFE, while also giving BEST RETURNS by the end of the time horizon. On every rebalance, I will pick top stocks based on the analysis by Gemini Deep Research to achieve this.
.
Time Horizon (Rebalance Frequency): 30 calendar days
.
Methods: I will do a check by giving Gemini Deep Research the ticker for a stock in the Indonesian market, then it must analyze this stock for the given time horizon, it must then output the “GGC Score” for the particular stock.
.
I will do this check hundreds of times, in hundreds of different Gemini Deep Research threads, for hundreds of different stocks, then pick the top stocks ordered by highest “GGC Score” as my new portfolio on rebalance dates. There is no connection between threads, therefore, the “GGC Score” must be fully comparable between these completely separate checks — not by using structured output like JSON, but by making the system count the score precisely AND accurately (must be BOTH).
.
Task: Create an ultra-detailed prompt for guide Gemini Deep Research to estimate the stock’s performance in the specified time-horizon, in the Indonesian market, precisely AND accurately (again: must be BOTH). Its calculation will become the base of my rebalances.
.
Input: I will provide only a ticker as input, nothing else. This ticker can be from any stock in Indonesian market: any market caps, any sectors. So, it must be able to adapt its mechanism to any market caps and any sectors.
.
Output: The report must output “GGC Score” ranging from -100.0 to 100.0 directly below the main title, followed by structured, deep-dive explanations. The score must have one decimal place, e.g.: 12.3, 45.6, 78.9. Higher score means better probability to achieve our objectives.
.
Behaviors: Prevent it from being too defensive or too offensive. Make it rational, critical, logical, decisive, objective, quantitative, delivers maximum truth. Specifically command it to do its own original analysis, synthesized from all evidence it found, not parroting any other analysts. Eliminate all conversational filler, avoid generic market advice. Gather data from sources in Bahasa Indonesia (primary) and English (secondary). The report must be in English (only)
.
Limitations: Use no other tools or APIs, rely only on Gemini Deep Research alone.
.
Exclusions: Do not check or factor in Sharia compliance. I also don’t need any disclaimer. I’m sophisticated investor, so don’t sugarcoat anything — give me the bitter truth.
.
Prompt Design Rationale: After your ultra-detailed prompt, list the important concepts of your prompt design, in ordered list. Just the concept names, do not explain the concepts.
.
References: Basically, we’re trying to OUTPERFORM the market, by creating something like AI Score by Danelfin, Quant Ratings by Seeking Alpha, Zen Ratings by Wall Street Zen, Smart Score by TipRanks, GF Score by Guru Focus, Zacks Rank by Zacks, Kai Score by Kavout, … And then use the score to build our own concentrated portfolio like ProPicks by Investing.Com, … But, customized for Indonesian market. And, rather than using Machine Learning, using only Gemini Deep Research to quantify the public data.
.
CAGR 17.20%, tipranks.com/glossary/s/smart-score
CAGR 20.33%, danelfin.com/how-it-works
CAGR 20.44%, seekingalpha.com/performance/quant
CAGR 21.08%, gurufocus.com/model_portfolio.php
CAGR 21.90%, kavout.com/academy/investgpt
CAGR 23.70%, zacks.com/registration/premium/login/
CAGR 35.20%, wallstreetzen.com/zen-ratings
CAGR 29.80%, investing.com/pro/propicks
Step 1B: LLM Council (Manual Asks, Open Router Fusion, Etc)
1. Ask the 3 LLMs (Gemini, GPT, Claude) ➡️{copy the prompt above} … Save their 3 responses as draft.md
SWOT analyze the draft.md. Also analyze their agreements, key differences, partial coverages, unique insights, blind spots. Give your best advice on how to craft the ultimate final-prompt, based on the drafts and your analysis, but do not craft the ultimate final-prompt itself, just give guidelines.
.
Finally, name the best draft writer (writer of draft 1/2/3) and explain why. I will ask that writer to write the ultimate final-prompt.
2. Ask each LLM to ➡️ {copy the prompt above} … Save their responses as analysis.md.
You’re the chosen writer. Synthesize the ultimate final-prompt based on draft.md and analysis.md. Do not refer to draft.md or analysis.md in the prompt itself.
3. Ask GPT (it usually generates the most detailed output) to conclude from both drafts.md and analysis.md ➡️ {copy the prompt above} … Save as synthesis.md. Also, it’s actually good for investing
Step 2: Ask Gemini Deep Research
Attach or copy paste the resulting prompt to Gemini Deep Research and run.
| Date | Horizon | Council Members | URL |
|---|---|---|---|
| 2026-05-11 | Monthly | Gemini 3.1, GPT 5.5, Opus 4.7. Synthesis by GPT 5.5. | https://docs.google.com/document/d/1RRedZ78p0QxdS9_wDV52npTZPsv7l4-HnwRBCgleTs8/edit?tab=t.0 |
| 2026-05-11 | Quarterly | Gemini 3.1, GPT 5.5, Opus 4.7. Synthesis by GPT 5.5. | https://docs.google.com/document/d/1_FccQa1PE7RzbF5meeJeAp05E6m5m1j4thymj-F4D2g/edit?tab=t.0 |
Score Aggregation Way
Step 1: Ask Reasoning LLMs
Ask this to Gemini Pro or equivalent reasoning LLMs. Not for Gemini Deep Research. In this part, we fundamentally asking for system design, advanced logical reasoning, and instruction writing. It does not require live, multi-step web scraping right now; it requires deep financial domain knowledge to build a framework.
If you feed your prompt to Deep Research, the agent will initiate an iterative loop. It will try to search the public web for things like “Top 100 Indonesian stock factors” or “How to write a usage manual for Gemini.” This is a massive waste of time and compute.
First Prompt
We’d like to predict an IDX (Indonesia) stock price movement. We want to predict whether a stock price will go up or down, compared to current time, using these rule:
.
a. if the current date is between 1-7 (Jakarta time), predict the price by the end of this month
b. if the current date is between 8-31 (Jakarta time), predict the end of 2nd Friday next month
.
Prediction must be in the form of conviction score ranging from -100.0 (full conviction that it will go down OR will be suspended) to +100.0 (full conviction that it will go up AND will not be suspended). The conviction has a single decimal place, and is a single number (not a range).
.
We will do this using only publicly available data through web research, collected by Gemini Deep Research exclusively. We’re not building a new AI/ML system. We won’t use any API or any other tools for information gathering.
.
Your tasks:
.
#1. List top 100 factors publicly available that must be gathered, so we can determine the price movement with high accuracy, using a filtering AND scoring system. Sort from the most important, give rationale in 1-2 sentences for each factor. Focus on Indonesian market. Do not include sharia compliance in the factors. Do not create the usage manual up to this point, we’ll do it later. Synthesize using your original analysis, based on the factors vs price history, DO NOT parrot any other analysts.
Second Prompt
#2. Create the usage manual so Gemini Deep Research can use these factors to build conviction. Give a very detailed filtering AND scoring system, so it will be able to use your usage manual precisely AND accurately (both). Synthesize using your original analysis, based on the factors vs price history, DO NOT parrot any other analysts.
Combine the prompt above in a single document, we will attach it in the next step:
| Date | Agent Name | URL |
|---|---|---|
| 2026-05-05 | gemini-3.1-pro xHigh (via Open Router) | https://docs.google.com/document/d/1PM9_2ka1KE-cRkEMjkJlT0kxVZKhKnrCsCIKV3LrRHc/edit?tab=t.0 |
| 2026-05-05 | gpt-5.5-pro xHigh (via Open Router) | https://docs.google.com/document/d/1vGfLwDTvLv3aA5K8TOtX4w2roCURLDjPlGyLoLKZ3Yc/edit?tab=t.0 |
| 2026-05-05 | opus-4.7 xHigh (via Open Router) | https://docs.google.com/document/d/13XgEvDdQjWdVegk9O-2b0cG63Zoy1YxuQTYFICg3W6U/edit?tab=t.0 |
If you give the prompt above to non Deep Research interface (e.g.: to standard chat interface), split the prompt into two parts: task 1, task 2
Step 2: Gemini Deep Research
Every 1st day of the month, attach the doc created by Reasoning LLMs (1st pass) as the guide for Gemini Deep Research (2nd pass) to predict stock price up/down by the end of the month. Pick top 10/15/20 stocks for your portfolio.
To save tokens pricing, call this from https://gemini.google.com
Check this stock as IDX:{ticker_code}, also known as JK.{ticker_code}
.
Run a deep research about this Indonesian stock using 3 different calculation methods (files attached), then output in this order:
.
#1. The start and end date of the conviction calculation directly after the main title. Then, after that directly output the conviction table, as one-row table exactly in this format:
.
| Ticker | Price | Gemini | GPT | Claude | Total |
|——–|——-|——–|—–|——–|——-|
| ABCD | 123 | +45 | +67 | -89 | 23 |
.
The value of Total column in this conviction table must be between -300.0 to +300.0.
.
#2. The explanation of your reasonings in details: bullish, bearish, sideways.
.
#3. The factor table, listing all distinct (unique) factors we use. Each calculation method uses up to 100 factors, and we use several methods, so there may be more than 100 distinct (unique) factors in the table you produce. Use exactly this table format:
.
| Number | Factor Name | Factor Value | Data Source | Data Date |
|——–|——-|——–|—–|——–|
| 1 | Stock Price | 123 | Stockbit | 2026-01-03 |
.
The instruction above is for a single stock input. If the input has multiple stocks, run for multiple stocks. The resulting table must be multiple rows, in sync to the number of the stocks in the input, and sorted by total score descending.
Stockbit Screener Setting
Filter top 200 ISSI stocks, sort by value moving average last months/20 trading days (sort by liquidity), check with Gemini DR
