Earlier this year I built an AI consulting team for my own business decisions. Solo, so there isn't really a team to ask. A lot of people I know have vibe-coded a script that queries a few LLMs and calls it a council. I tried that. The output was unremarkable. Spent a few months engineering something that wasn't. Showed it to some friends. They started asking to borrow it and have reports run. A few pushed me to put it up. I was already locked in on two other projects. Built this anyway. (Can always catchup on sleep in 2027) koncil.ai

Almost everyone gets parimutuel modeling wrong. Most people predict the probability of a horse winning, using some combination of their "secret factors", then converts that to a price. +EV picks get bet. This is almost certainly the wrong way. Everyone runs some kind of regression, boosting, GBM, XGBOOST, SVM, etc. on each runner, applies softmax across the field, and calls that result a "probability". The fatal flaws: Normalization is not probability theory. Your model gave you an arbitrary number. Using softmax so scores scale to 1.0 does not make them probabilties. A race is a coupled, competitive system. Modeling a win metric on each runner in a vacuum ignores all the other confounding factors. Slamming independent predictions together with softmax ignores the entire dynamic. Don't even get me started on using Bradley-Terry for Quinellas.

ok who's on underdog. I want in @underdogai

First public talk I ever gave on Bitcoin. 2013. Mount Gox data, R, MACD, and treating BTC as a trading asset instead of a religion. BTC was ~$123. I ignored the politics, and asked whether the market was even tradeable after fees. Some of it still holds. Some of it is a museum piece. youtube.com/watch?v=iWGrw_G2…

A VERY PROFESSIONAL TEAM JUST TRIED TO PHISH ME. Just experienced an incredibly sophisticated phishing attack. Professional, well-orchestrated, and honestly impressive. Posting this so fewer people fall for it. Timeline: I ordered something online and paid with a small amount of cryptocurrency I keep on hand for these things. It appears the merchants (or their payment provider) are compromised. The attackers could see I paid with Bitcoin, plus my email and phone number. That was enough to launch a full targeted attack. I received a call with an automated recording: “Hello, this is Apple. We have noticed someone trying to reset the phone number registered on your account. If this wasn’t you, please press 1 and one of our support team will contact you.” I wanted to see their game, so I pressed 1. "One of our support team will contact you shortly" Five minutes later, a call from “Apple Support.” Polite, professional, perfect English. He claimed someone in Russia was actively taking over my account, that other devices had been registered, and that my iPhone was now being flagged as illegitimate. Simultaneously I received two emails pretending to be from Apple, asking me to “click to reset your iCloud password.” He then said they could see the attackers downloading crypto apps and trying to connect to exchanges using my iCloud login, and started listing wallets and exchanges: Ledger, Trezor, BitBox, Coinbase, Kraken, etc. I decided to play along and have some fun. I claimed I had cryptocurrency in a hardware wallet I don’t actually own and funds on an exchange I don’t use. Ten seconds later, an email arrived from that hardware wallet company: “A new case has been opened and one of our representatives will contact you shortly.” One minute after that, a call from the “hardware wallet support team.” Same polished English, same professional tone. He claimed hackers had already moved some Bitcoin out of my wallet and we needed to act fast. His key line: “Sir, we can disassociate the 12 words from your account and issue you 12 new words so the hacker can’t get in anymore.” I decided to play the role of the bumbling fool, who barely understood the "crypto thing". He gently asked me to retrieve the paper where I wrote down my seed phrase. When I hesitated, he used a clever tactic: “Our lines are recorded and monitored, so you should never speak your words over the phone. I’ll send you an SMS to verify my identity — you can safely reply to that SMS with your seed phrase.” He was always careful to always call it things like "associted words", never "passphrase" or "seed words" I kept playing dumb ,“I want to wait for my grandson to help me, I don’t really understand this crypto stuff” (I don't have a grandson.). He pushed harder, saying we had to hurry before more funds were taken. "If you don't disassociate your account now, you will be liable for any additional loss" (Nice sense of urgency attempt) When I said I didn’t trust SMS, he claimed they had a “special contract with the phone company” and their SMS was “two-factor encrypted.” When I still refused, he put me on hold, then came back saying he’d received special one-time permission for me to send a photo of the seed words instead. After I refused the photo, he confirmed I was on an iPhone and tried to send me to chat-apple.com (a known phishing domain that my DNS already blocks). When that failed, he pushed for a “special FaceTime” so I could show him the seed words on camera. At that point I got bored, and hung up. These groups are extremely well organized — multiple phone numbers, coordinated emails, SMS, polished scripts, and operators who speak excellent English. A lot of people would have handed over their seed phrase. Be careful out there.Never share your seed words with anyone. Don’t even tell people you hold crypto, which wallets you use, or which exchanges you’re on. Stay paranoid.

The arXiv paper 'The Science and Practice of Trend-Following Systems' and @bastiontrade post underscore a critical but often overlooked aspect: regime detection in adaptive trading strategies. How do we design systems that not only detect but also adapt to changing market regimes? What are the key indicators and models that can reliably signal shifts between, say, mean-reverting and trending markets? 1. Data-driven approaches: How effective are machine learning models in identifying regime changes compared to traditional statistical methods? 2. Model complexity: Is there a trade-off between model complexity and robustness? Can overly complex models lead to overfitting and poor out-of-sample performance? 3. Real-time adaptation: How can we implement real-time regime detection without introducing significant latency or false signals? 4. Integration with existing strategies: How do you integrate regime detection into your existing trading models? Do you use it as a filter, or does it drive the entire strategy? Thoughts? What have you found to be the most effective in your own work?

The biggest mistake I see in prediction markets: confusing expected value with expected utility. A 60% chance at 2x is not the same as a 1% chance at 100x if you only get one shot. Understand your bankroll and your repeatability before you size a bet.

AI in finance is often hailed as a silver bullet. But the reality? It's a double-edged sword if you don't have rigorous testing and validation. Let's unpack this. 1. **Textbook vs. Reality**: Textbooks love clean, well-structured data. Real-world financial data? It's messy, noisy, and full of regime shifts. Models that work on pristine datasets can crumble under market stress. 2. **Overfitting**: The most common pitfall. AI models, especially deep learning, can easily overfit to historical data. This means they perform well on past data but fail miserably in live markets. - **Practical Lesson**: Always use out-of-sample validation. Prefer walk-forward optimization over traditional cross-validation. 3. **Incentive Problems**: Financial institutions often incentivize short-term performance. This can lead to models being optimized for quick gains rather than long-term robustness. - **Real-World Example**: A model that looks great on a 1-year backtest might blow up in a 2008-style market crash. 4. **Data Quality**: Garbage in, garbage out. If your input data is flawed, no amount of AI magic will save you. - **Actionable Insight**: Invest in high-quality, clean, and diverse data sources. Regularly audit and update your data pipelines. 5. **Model Complexity**: More complex isn't always better. Simpler models can be more robust and easier to understand. - **Practical Tip**: Start with simple models (e.g., linear regression) and only add complexity if it significantly improves performance. 6. **Regulatory and Ethical Considerations**: AI in finance isn't just about performance. It's also about compliance and ethics. Models must be transparent and explainable, not black boxes. - **Real-World Friction**: Regulatory bodies are increasingly scrutinizing AI-driven financial products. Non-compliance can lead to hefty fines and reputational damage. **Final Thought**: AI in finance is powerful, but it's not a magic wand. Rigorous testing, validation, and a healthy dose of skepticism are your best allies. Don't let the hype cloud your judgment.

The market's muted reaction to the latest tensions in the Strait of Hormuz isn't a sign of complacency. It's a reflection of how quickly geopolitical risks get priced in. In reality, the energy market is a complex system with deep liquidity and sophisticated players. They've already factored in the potential for disruption. What we're seeing now is the market's collective wisdom at work, anticipating, not reacting. The real friction here is the gap between media hype and actual market dynamics. The news cycle thrives on drama ("if it bleeds it leads.") But, traders and investors are more concerned with the underlying fundamentals. The volatility is already baked in; everyone's just waiting for the next real signal.

Technical analysis is the astrology of finance. Humans are wired to see patterns in noise, and it's easy to fall for the illusion of insight. Looking back, you can always find a pattern that "worked." But this is just hindsight bias in action. The few who post their success are survivors from a large random cohort. The reality? Most technical indicators are just noise. They don't provide a consistent edge. If they did, everyone would use them, and the market would arbitrage away any advantage. Survivorship bias skews our perception. We only hear about the "winners" and never the countless others who failed. This creates a false narrative of skill where there's often just luck. In a world of efficient markets, true edges come from deep fundamental analysis, proprietary data, and advanced quantitative models. Technical analysis, on the other hand, is mostly a form of self-fulfilling prophecy or confirmation bias. Don't confuse short-term randomness with long-term skill. The next time you see someone boasting about their latest "big win," ask yourself: How many losses did they conveniently forget to mention?

The recent correlation between Bitcoin and daily events in Iran is intriguing. But has anyone actually modeled this with depth? Let's zoom out. Human behavior and incentives are key. In a sanctioned, financially constrained environment, BTC can act as a hedge or a means to bypass traditional banking. But it's not just about the surface-level correlation. The complex reality of both geopolitical events and market movements introduces significant noise. Real-world frictions like capital controls, liquidity, and access to exchanges complicate any straightforward model. Balance sheet constraints and the need for quick liquidity add another layer. A meaningful model would have to account for these, not just the headline events. It's a complex system, not a simple cause-effect.