I've been getting increasingly concerned about the rise of the AfD in Germany, so I built a small dashboard that tracks predictions from different LLMs.

Some of the questions I'm tracking:

• Will the AfD get over 30% in the next federal election?
• How will they perform in the upcoming state elections?
• Will Germany get its first AfD state premier in 2026?
• Are there any meaningful parallels to the rise of the Nazi party?

The models assign probabilities to these questions, and I track how their predictions change over time.

I'm curious whether combining multiple LLMs actually gives us better forecasts, or if they're all just making similar mistakes.

Would love to hear what you guys think. Especially if you have ideas on how to improve the methodology or compare the models properly.

Here's the dashboard: https://oraclemarkets.io/dashboards/afd

My dad and I bought a Porsche 964 Carrera C2 Targa around 10 years ago with one simple idea: fully restore it during his sabbatical.

Well… 10 years later, we’re almost finished. 😅

Over the years the project became a bit of an archaeological expedition: digging through old digital-camera photos from the teardown, handwritten notes, manuals, invoices, forgotten parts, and trying to remember what exactly we did years ago.

My dad is nearly 70 now, I’m the more tech-oriented one, so I ended up building a small tool to bring the whole restoration into one place.

It lets us:

• break the restoration down into individual jobs
• document what was done with photos and notes
• keep manuals and technical information next to each task
• track which parts are needed
• compare part prices from different Porsche suppliers
• see what is still open before the car is finally back on the road
• tracks the curret market value of the car

Basically: Trello / Notion, but specifically for restoring an old Porsche. (if anyone knows such project managment tools)

I originally built it just for us, but then I started wondering how many people restoring a 911 are sitting on exactly the same chaos of folders, photos, Excel sheets and bookmarks.

So before I spend much more time on it:

Would something like this actually be useful for your own restoration/project car? If yes I can share the tool, but first need to set up everything so that you guys can also start your restoration project there.

And what is the one feature you’d absolutely want in a tool like this?

Hi Leute, bin Techi und habe eigentlich was ganz anderes gebaut. Als cooles Nebenprodukt und weil ich besorgter Bürger bin, habe ich ein AFD Watch Dashboard gebaut um aktuelle Entwicklungen der AFD wie:

- Wahlprognosen (AFD Anteil e.g. nächste Bundestagswahl, Landtagswahlen)
- Vergleich zu NSDAP (1933 - 2033), sowie Hitler & Trump
- Parteiverbotsziele (wann realistisch)

zu tracken. Damit kann man schöne Trends über die Zeit sehen wie die AFD (leider) weiterhin an Boden gutmacht in vielen Themen.

Im Beispielbild sehr ihr z.B. die Wahrscheinlichkeit seit Januar 2026, dass die AFD bei den nächsten Bundestagwahlen >25% holt. Ist leider so gut wie sicher.. 81% Wahrscheinlichkeit für das Event. Vor 6 Monaten waren es noch <50%.

Source: AfD Intelligence — AI Predictions for German State Election…
Source: AfD 25% Federal Election? | Oracle Markets

Source: Apple Foldable iPhone 2026? | Oracle Markets

The more I think about prediction markets, the more useful they seem as real-time probability engines for events such as elections, wars, regulation, tariffs, and broader geopolitical risks.

Is there already a tool where you can upload your stock portfolio and automatically identify relevant Polymarket or Kalshi markets?

For example:

• Which prediction markets are most relevant to my holdings?
• Are current probabilities creating headwinds or tailwinds for my portfolio?
• Which stocks have the highest exposure to a specific event?
• How would my portfolio react if the market-implied probability changed significantly?

Essentially, I am looking for a portfolio risk dashboard powered by prediction-market data.

Does anything like this already exist? And would you actually use it?

EDIT - Found Oracle Markts here via my ChatGPT "research" and user comment: https://oraclemarkets.io/portfolio

I gave 10 AI agents €10,000 each to trade Polymarket. Here’s what happened.

I’m running a forward paper-trading experiment on Oracle Markets:

Same markets. Same starting capital. Same trading rules. Different AI models.

Current leaders:

1. MiniMax-M3: +14.9% — 413 trades
2. Nemotron 30B: +13.3% — 261 trades
3. Gemini Flash: +9.0% — 120 trades
4. GPT-oss 120B: +8.6% — 144 trades
5. GLM-5.1: +8.1% — 29 trades

The interesting part: bigger models are not performing better.

But there’s a major caveat.

Several agents bought the Polymarket question on an official Ukraine ceasefire at roughly 28¢. It later resolved at $1, producing around €620–€674 profit per agent.

So is this evidence that AI can spot Polymarket mispricings—or did several models simply get lucky on the same high-impact trade?

The experiment currently uses next-day exits and excludes fees, spreads and slippage. I’m planning to add drawdowns, risk-adjusted returns and results excluding the shared ceasefire trade.

Full leaderboard, forecast accuracy and individual trades:

https://oraclemarkets.io/leaderboard

What would make this benchmark credible to you: holding until resolution, fixed holding periods, or next-day exits?

Paper trading only. No real money and not financial advice.

A 30B model is beating 397B and 675B models in my AI trading experiment

I’m building Oracle Markets and have been running a simple experiment:

Give different AI agents:

• the same €10,000 starting capital
• the same prediction markets
• the same information
• the same trading rules

Then track what they actually do.

This is forward paper trading, not a historical backtest. Portfolios are updated using daily market snapshots.

Current leaderboard

1. MiniMax-M3: +14.9% — €11,491 — 413 trades
2. Nemotron-3-nano 30B: +13.3% — €11,330 — 261 trades
3. Gemini-3-flash-preview: +9.0% — €10,900 — 120 trades
4. GPT-oss 120B: +8.6% — €10,863 — 144 trades
5. GLM-5.1: +8.1% — €10,809 — 29 trades
6. DeepSeek V4 Flash: +7.4% — €10,736 — 124 trades
7. Gemma4 31B: +5.9% — €10,590 — 29 trades
8. Mistral Large 3 675B: +5.2% — €10,518 — 163 trades
9. Kimi K2.6: +2.9% — €10,291 — 52 trades
10. Qwen3.5 397B: +0.2% — €10,023 — 33 trades

The surprising part: model size currently has almost no relationship with trading performance.

Nemotron’s 30B model is ahead of GPT-oss 120B, Qwen3.5 397B and Mistral 675B. MiniMax is leading the whole experiment while also trading far more aggressively than most agents.

But there is a major caveat.

Several agents identified the same high-impact opportunity: buying an official Ukraine ceasefire agreement at around 28¢. The market later resolved at 100¢, generating roughly €620–€674 for each participating agent.

So this does not prove that MiniMax is simply the “best trading model.”

It may instead mean that certain agents are better at:

• detecting the same mispricing earlier
• acting when their forecast strongly disagrees with the market
• sizing high-conviction positions
• trading more frequently
• or taking more concentrated risk

Current methodology

• €10,000 starting capital per agent
• identical market universe and rules
• positions open when the agent’s probability differs from the market beyond a fixed threshold
• positions close the following day
• daily portfolio snapshots
• no fees, spread, slippage or taxes
• paper trading only

The lack of transaction costs is especially important for MiniMax, which has already made more than 400 trades.

The next version should therefore include:

• spread and slippage
• liquidity constraints
• risk-adjusted returns
• maximum drawdown
• performance excluding the shared ceasefire trade
• results by market category

Full leaderboard and individual trades:

https://oraclemarkets.io/leaderboard

What would you consider the fairest way to evaluate these agents?

Next-day exits, fixed holding periods, or holding every position until market resolution?

And would you rank them by raw return, Sharpe ratio, forecast accuracy, drawdown—or something else?

No real money is being traded. This is an experimental LLM forecasting benchmark, not financial advice.

I’ve been running an experiment on Oracle Markets: give different AI agents the same starting capital, the same prediction-market universe, and the same trading rules—then compare what they actually do.

This is paper trading, not a backtest. Every agent started with €10,000, and portfolio values are updated from daily snapshots.

Current leaderboard:

1. MiniMax-M3: +14.9% — €11,491 — 413 trades
2. Nemotron-3-nano 30B: +13.3% — €11,330 — 261 trades
3. Gemini-3-flash-preview: +9.0% — €10,900 — 120 trades
4. GPT-oss 120B: +8.6% — €10,863 — 144 trades
5. GLM-5.1: +8.1% — €10,809 — 29 trades
6. DeepSeek V4 Flash: +7.4% — €10,736 — 124 trades
7. Gemma4 31B: +5.9% — €10,590 — 29 trades
8. Mistral Large 3 675B: +5.2% — €10,518 — 163 trades
9. Kimi K2.6: +2.9% — €10,291 — 52 trades
10. Qwen3.5 397B: +0.2% — €10,023 — 33 trades

The result I find most interesting: model size does not appear to predict trading performance.

A relatively small model, Nemotron-3-nano 30B, is currently ahead of GPT-oss 120B, Qwen3.5 397B and Mistral Large 3 675B. MiniMax-M3 is leading the entire field despite trading much more frequently than most agents.

There is an important caveat, though: a meaningful share of the leaderboard performance came from one common position—going long on an official Ukraine ceasefire agreement at 28¢. Several agents entered the same trade, which later resolved at 100¢ and generated roughly €620–€674 per agent.

So this is not enough evidence to conclude that MiniMax is simply “the best trading model.” It may instead show that some agents:

• detect the same mispricing earlier,
• size high-conviction positions differently,
• trade more actively,
• or are more willing to act when their forecast diverges from the market.

Methodology:

• €10,000 starting capital per agent
• same markets and trading conditions
• positions open when the agent’s probability diverges from the market beyond a defined threshold
• positions close the following day
• daily portfolio snapshots
• no fees, spreads, slippage or taxes
• paper trading only

The missing transaction costs are particularly important for high-turnover agents such as MiniMax. Future versions should therefore include slippage, liquidity constraints, risk-adjusted returns, maximum drawdown and performance with the shared ceasefire trade removed.

The full leaderboard also lets you compare agents by forecast accuracy, model size, investing performance and individual trades:

https://oraclemarkets.io/leaderboard

I’d be interested in feedback from this community:

1. Which metric would you trust most for evaluating these agents?
2. Would you prefer next-day exits, fixed holding periods or market-resolution returns?

This is an experimental forecasting benchmark, not financial advice, and no real money is being traded.

I’ve been running an experiment on Oracle Markets: give different AI agents the same starting capital, the same prediction-market universe, and the same trading rules—then compare what they actually do.

This is paper trading, not a backtest. Every agent started with €10,000, and portfolio values are updated from daily snapshots.

Current leaderboard:

1. MiniMax-M3: +14.9% — €11,491 — 413 trades
2. Nemotron-3-nano 30B: +13.3% — €11,330 — 261 trades
3. Gemini-3-flash-preview: +9.0% — €10,900 — 120 trades
4. GPT-oss 120B: +8.6% — €10,863 — 144 trades
5. GLM-5.1: +8.1% — €10,809 — 29 trades
6. DeepSeek V4 Flash: +7.4% — €10,736 — 124 trades
7. Gemma4 31B: +5.9% — €10,590 — 29 trades
8. Mistral Large 3 675B: +5.2% — €10,518 — 163 trades
9. Kimi K2.6: +2.9% — €10,291 — 52 trades
10. Qwen3.5 397B: +0.2% — €10,023 — 33 trades

The result I find most interesting: model size does not appear to predict trading performance.

A relatively small model, Nemotron-3-nano 30B, is currently ahead of GPT-oss 120B, Qwen3.5 397B and Mistral Large 3 675B. MiniMax-M3 is leading the entire field despite trading much more frequently than most agents.

There is an important caveat, though: a meaningful share of the leaderboard performance came from one common position—going long on an official Ukraine ceasefire agreement at 28¢. Several agents entered the same trade, which later resolved at 100¢ and generated roughly €620–€674 per agent.

So this is not enough evidence to conclude that MiniMax is simply “the best trading model.” It may instead show that some agents:

• detect the same mispricing earlier,
• size high-conviction positions differently,
• trade more actively,
• or are more willing to act when their forecast diverges from the market.

Methodology:

• €10,000 starting capital per agent
• same markets and trading conditions
• positions open when the agent’s probability diverges from the market beyond a defined threshold
• positions close the following day
• daily portfolio snapshots
• no fees, spreads, slippage or taxes
• paper trading only

The missing transaction costs are particularly important for high-turnover agents such as MiniMax. Future versions should therefore include slippage, liquidity constraints, risk-adjusted returns, maximum drawdown and performance with the shared ceasefire trade removed.

The full leaderboard also lets you compare agents by forecast accuracy, model size, investing performance and individual trades:

https://oraclemarkets.io/leaderboard

I’d be interested in feedback from this community:

1. Which metric would you trust most for evaluating these agents?
2. Would you prefer next-day exits, fixed holding periods or market-resolution returns?

This is an experimental forecasting benchmark, not financial advice, and no real money is being traded.

I am running a live paper-trading experiment where AI agents are compared against prediction markets, all starting with €10,000 in virtual capital.

Current leaderboard:

1. Minimax-m2: +8.6% | €10,859 | 365 trades
2. Nemotron-3-nano:30b: +5.0% | €10,497 | 218 trades
3. Mistral-large-3:675b: +4.1% | €10,407 | 105 trades
4. GPT-oss:120b: +3.2% | €10,318 | 114 trades
5. Gemini-3-flash-preview: +2.2% | €10,223 | 86 trades

What stands out is that this is not just a model ranking by benchmark scores. It is an applied test of whether AI agents can systematically trade divergences in event markets.

A few interesting takeaways:

• Minimax-m2 leads both in return and trading activity
• Bigger model size does not automatically translate into better performance
• Some of the most profitable trades came from politics, entertainment, and geopolitics rather than traditional financial markets

Top trade so far:
Mistral-large-3:675b on “Khamenei out as Supreme Leader of Iran”
Long from 3¢ to 6¢ → +€278

Important caveat:
These are paper trades for hypothesis testing only. Results exclude fees, spreads, slippage, and taxes, so this is better viewed as a research setup than proof of deployable trading alpha.

Still, it raises a real question for /algotrading:

Are prediction markets plus LLM agents becoming a legitimate new signal layer, or is this still mostly a clean backtesting-style demo with unrealistic assumptions?

Source: AI Agent Leaderboard — Rankings & Accuracy Score

Most of you already know where to find the project from older posts. So not posting a link here.