400 days of Crash Brief

Today it has been exactly 400 days since I started Crash Brief, a crypto risk assessment service for B2B, especially for people working with options. I want to look back a little at the stages I went through, what worked, where I was wrong and what I eventually arrived at.

From the beginning, I was betting on AI development. I thought more powerful models would appear, and with them new ways to analyze the market and make more accurate forecasts. But over this time, almost nothing fundamental has changed for me with forecasting.

Here is the thing. You find a model for looking at the market, it shows an interesting picture in backtests, it works for a while. Then different participants enter the market, conditions change, and the same approach stops giving the previous result. And you have already had time to think that you finally found something.

This was probably my main miscalculation, expecting that more powerful AI would let me reach stable forecasting accuracy. I have not achieved a consistent 70% so far. I have gone through many approaches, analyses and hypotheses known to me, and everything turns out to be very unstable. Now I increasingly think the forecasting approaches themselves will have to change systematically, because the market we are trying to forecast keeps changing.

I started with Ethereum alone. The deeper I analyzed it, the more I understood that I could not really do without Bitcoin. Its chart has its own clues, there is more liquidity and money there moving the whole market. Gradually I added BTC, ETH, SOL and HYPE, and then dropped HYPE.

From my observations, markets with a strong institutional layer and markets where it is weaker behave differently. You cannot just take the same approach and expect it to work equally on every asset. For HYPE, I would have had to develop a lot from scratch, and I understood that this was already taking away my focus. I put it aside.

A separate story is how much time went into autonomy. Data collection, assessments, forecasts, many different indicators, everything came together piece by piece. And practically every day gave feedback. Something broke somewhere, sometimes I did something wrong, sometimes the AI did. Sometimes an error appears immediately, sometimes a week or a month later. Plus exchanges, load, broken connections, new functions that I keep adding.

At the start I looked at options, open interest and some basic things. Now I can examine what is happening in a specific zone in much more detail. How many liquidations there were, how they were distributed across exchanges, what is happening in spot and futures, how order book orders change as the price approaches. Orders appear, disappear, appear again. This gives additional observations about participants' behavior, although of course we do not know their actual intentions.

And this is where I really felt the value of each day. A day passes, outcomes appear, you can draw conclusions and move on. However many models and backtests I build, tomorrow will happen only once. I cannot speed that up. And regularly the market turns out to behave in a way I did not expect at all.

A large separate area of work developed around options economics. You have an asset and want to buy a protective option in case it falls. How much does this insurance cost, what will it cover and under which outcome? At one moment it is expensive, at another the conditions may be more interesting. On the other side are the sellers of this insurance, risking their capital and also assessing every day how much risk they are taking.

Right now I have a main approach that I continue to use. I do not consider it a holy grail, but so far I have not been able to offer a convincingly better replacement. I make manual adjustments, observe the result and examine the errors.

In parallel, I am building an automatic version into which I am trying to transfer this working process itself. Get an outcome, look at where I underestimated or overestimated risk, ask questions, examine possible reasons, think of what else to test. There can be a lot of questions. A model can spend a long time examining them and suggesting new hypotheses, but then they still have to be tested on the following days.

I want to give this version 90 days and see whether it can assess risk, volatility and options economics better than I do in a semi-automatic mode. Organizing all of this took many hours. Something still breaks from time to time, so I have not started the countdown yet. It is close, but it is too early to announce a fully ready autonomous system.

Autonomous work is possible, I see it in practice. At the same time, organizing this work feels like architectural creativity to me. You need to set goals, conditions and boundaries so that a powerful model keeps moving where you want it to go. Sitting next to it and pushing it with every next prompt gets tiring, and there is simply no time for it.

We will see what happens over these 90 days. It may turn out that the automatic version does not give an advantage either. That would also be a result, although of course I would like something else.

At some point I was very frustrated. So many years in the market, so many methods tried, and I still cannot get the forecasting accuracy I imagined. In retrospect everything can look very beautiful. You select an approach, look at the history, it seems to work. Whether it will work next week is another question entirely.

But gradually another interesting perspective appeared. I look at job openings at funds and see them searching for people who are supposed to build something similar to what I am working on. Smart people, developers, people with a financial background who know how to work with AI. And I think many of them will encounter the same difficulties.

I do not know how every fund works and cannot speak for each of them. But now I am much more careful with the idea that somebody must have a miracle model that predicts the future well. In results, it can be difficult to separate the quality of an approach from luck. Two teams may use serious methods, one gets lucky, the other does not. The first decides it found the right way, the second looks for where it went wrong.

I am trying to understand better which force is driving the market now and which approach even makes sense in these conditions. Today it is one driver, tomorrow another, then they combine and change each other. This is the combination I have to work through.

And more and more of my attention goes to the economics of a particular decision. Is this trade expensive or cheap right now? Does our particular portfolio need it? How much risk are we already carrying and how much more can we take? What would buying protection give us and how much would we have to pay for it?

The forecast remains part of this work, but portfolio and position management are making more and more sense to me. This is roughly where I am moving now.

400 days have passed, the experience has been tremendous. I now look at many conversations about crypto and trading very differently, because I have had time to test some of my assumptions and continue testing them. I hope something interesting will still turn up.