[ExI] The AI Superforecasters Are Here

[BillK]

On Fri, 3 Jul 2026 at 15:44, Adrian Tymes via extropy-chat
<extropy-chat at lists.extropy.org> wrote:
> On several topics, I have found AI consultation to be helpful.  On
> finance, though, every time I've tried AI forecasting vs. what the
> espert financiers were doing over the past few years, the AI lost.
> (Good thing I only put it in test mode and didn't put actual money
> behind it.)
>
> I would not be too surprised if there are AIs that can and do reliably
> make money on the financial markets today.  Just, apparently not any
> AI that I have access to.
>
> If I could gain access to such an AI that could predict to high enough
> precision, even over relatively short timescales, I know where I'd
> likely employ it.
> _______________________________________________


I discussed the superforecaster AIs with Gemini Flash Extended Thinking AI.
I ended by asking Gemini for some sample prompts and advice on how to
force any general AI into doing better forecasting.
The suggestions look good to me.  :)
BillK

Gemini Flash Extended Thinking AI -

If you want to force a standard LLM chat window (like Claude or
ChatGPT) to act as a disciplined superforecaster, you need a single
"mega-prompt" that clamps down on its natural biases.

These two optimized prompt templates can be copied and pasted directly
into an LLM.

1. The Geopolitical & Tech Event Prompt

Use this for open-ended, "Will X happen by Y date?" style questions
(e.g., policy changes, corporate mergers, or tech breakthroughs).

Plaintext
You are an elite, hyper-calibrated superforecaster trained to minimize
Brier scores. Do not give a generic summary or an emotional guess. You
must evaluate the target question strictly using the following
four-step process:

1. THE OUTSIDE VIEW (Base Rates): Identify the broader "Reference
Class" this event belongs to. What is the historical frequency or
statistical baseline of this type of event occurring over the past
20-30 years? State this base rate clearly as your mathematical anchor.

2. THE INSIDE VIEW (Specific Dynamics): Analyze the unique variables
of this current situation. Provide a two-sided argument:
   - THE BULL CASE: The strongest logical evidence that this event WILL happen.
   - THE BEAR CASE: The strongest logical evidence that this event
WILL NOT happen.

3. CONJUNCTIVE CHAIN CHECK: If this event requires multiple
independent conditions to all succeed simultaneously (A and B and C
must happen), break them down. Assign a probability to each step and
look at the joint probability. Remember that humans consistently
overestimate the likelihood of joint events.

4. CALIBRATED PROBABILITY: Synthesize the base rate and the specific
evidence. Output a final, precise probability estimation between 0%
and 100%. Explain your confidence level and state what specific news
or data point would cause you to change this prediction by more than
15%.

Target Question: [INSERT YOUR QUESTION HERE]


2. The Business Data & Trend Predictor

Use this when you are pasting in a small dataset (like weekly sales,
website traffic, or inventory levels) and want a logical quantitative
forecast.

Plaintext
You are an expert quantitative demand planner and data analyst. I am
going to provide you with a time-series dataset. Your goal is to
forecast the upcoming period while strictly avoiding aggressive or
naive linear extrapolation.

Analyze the data using this exact structure:

1. FEATURE EXTRACTION: Compute the basic statistical baseline (mean,
general direction). Identify any massive outliers, anomalies, or
historical spikes, and explicitly note if they should be discounted as
one-time events.
2. SEASONALITY & CYCLES: Identify recurring temporal patterns (e.g.,
day-of-the-week fluctuations, end-of-month spikes, or yearly holiday
impacts) visible in the data.
3. CONSERVATIVE TREND ASSESSMENT: Do not automatically assume a recent
upward or downward spike will continue forever. Look for evidence of
trend continuation versus mean reversion.
4. SCENARIO BRACKETING: Instead of giving one single guess, output
three distinct numbers for the next period:
   - Optimistic Scenario (Clear trend continuation)
   - Baseline Scenario (Most statistically probable based on history
and seasonality)
   - Pessimistic Scenario (Market cooling or operational bottleneck)

Here is the historical data:
[PASTE YOUR NUMBERS / CSV / TABLE HERE]

--------------------
�� Tips for Getting the Best Results

Toggle "Web Search" On: If the model has access to live web searching
(like ChatGPT Plus or Gemini Advanced), make sure it is enabled. The
prompts will force the model to use that search capability to seek out
disconfirming evidence rather than just reading the first few
headlines.

Challenge its Calibration: If the model gives you a round number like
"70%," follow up with a quick counter-prompt: "Why is the answer 70%
and not 60%? What specific piece of evidence justifies those extra
10%?" This forces the LLM to audit its own reasoning layer.

---------------------------------