The short version: two observations, one product. The first observation is that almost every retail AI trading tool puts the LLM in the wrong layer of the stack — at execution instead of translation — and that is where the surprise losses come from. The second observation is that almost every trading tool, AI or not, has no relationship with the user — and that is why most people open one for a week and then never come back. TraderBear is the bet that fixing both, at the same time, makes a category that didn't exist.
A friend asked Claude to help him "watch some stocks and trade them when momentum confirms." Claude wrote some Python, plugged in his broker API key, ran it on his laptop. Two weeks later he was down about $1,400 on what he thought was a paper-money test, because somewhere in the code path the agent had passed a flag he didn't notice and started using his real account. The script worked. The architecture didn't.
I saw a version of this story enough times to stop dismissing it. The pattern was always the same: the user wanted something they could describe in English; the LLM was happy to write code that did approximately that; nobody in the loop was responsible for the boring engineering questions — paper vs live, risk caps, audit logs, what happens when the API rate-limits you mid-order. The LLM is good at the creative parts. The boring parts are where the money lives.
The architectural bet behind TraderBear is that the right place for an LLM in a trading tool is at the rule-translation layer, not at the execution layer. You describe what you want in English. The LLM converts it into a structured rule. Code — not a language model — reads that rule, watches the market, enforces risk caps, and executes.
The benefit isn't theoretical. It's that I can answer "what stops the agent from doing X?" with a code path I can show you, not with "the system prompt says not to." A risk cap that lives in a prompt is a suggestion. A risk cap that lives in a Cloudflare Worker validating every order before it leaves is a cap.
I noticed something else during the same period. People who tried ChatGPT-for-trading were excited in week one, repeating themselves by week three, and gone by month two. Not because the model got worse — the model didn't change. Because each conversation was a fresh start. Re-explain your watchlist. Re-state your risk tolerance. Re-paste the rule you've been refining. The work of restating context is so much friction that the gains never compound.
Meanwhile, anyone who has ever raised a virtual pet (Tamagotchi, Neopets, Pou, whatever the current generation calls them) knows that persistent identity changes behavior in ways stateless tools cannot. You don't abandon a thing you've named. You check on it. You read the log of what it did. You don't override the routine you spent three weeks building, because changing it feels like undoing your own work.
The second bet behind TraderBear is that combining the two — an AI trading agent's capability with a virtual pet's persistent identity — is a category. You adopt a bear. The bear has a name you chose, a stage that evolves with use, an avatar that grows up alongside you. The trading engine underneath is conventional: plain English in, structured rules out, paper-first by default, risk caps in code. The pet frame is what makes you actually run the discipline.
An AI trading agent you adopt. It works across stocks, ETFs, crypto, futures, and prediction markets. You give it plain-English rules; it executes under risk caps you set. Paper money is on by default. The bear has stages — cub, apprentice, trader, partner — that gate capability as the user develops the discipline to use them. Every trade the bear takes is logged with the rule that fired and the inputs observed.
If you've used a retail AI trading tool before, the trading engine will feel familiar. The bear part is the part you can't replicate with a fresh ChatGPT session. By month two of using TraderBear, you've trained a specific bear on specific rules across specific categories, and the cost of starting over with another tool is high enough that you don't.
It is not a quant fund. The market math is not novel; the architecture and the pet are. There are no performance promises here — and there can't be, because TraderBear is not a registered investment adviser. What it offers is a tool with two properties: the safety architecture beneath the agent is enforced in code, and the relationship layer keeps you engaged with the discipline long enough for that architecture to matter.
It is not "AI trading bot for beginners" in the marketing sense — the segment where the implicit promise is "let the AI make money for you while you sleep." That promise is false on its face and we won't make it. The audience we want is the population that finds trading interesting, wants to learn, and would benefit from a tool that makes the learning enjoyable rather than punishing.
It is not built around any single asset class. The bear works across categories. Some users will run a single rule on a single market for months and never expand; others will keep adding rules across stocks, crypto, futures, and prediction markets. The bear's stage and learning generalize across all of them.
I'd like to be wrong about the size of the audience that wants a tool with both safety architecture and a relationship layer. I think it's bigger than the indie-hacker / quant Twitter crowd thinks. I think there is a real population of curious beginners and patient intermediate traders who would happily use a careful, gamified tool — and who currently default to nothing, or worse, to tools that promise the moon.
I'd also like to be wrong about how long the pet retention effect actually lasts. The hypothesis is that the bear's persistent identity adds months to user retention compared to a stateless tool. I'll know in 6-12 months whether that's true.
Three things, in order. Better backtests across more categories so the paper sessions are as honest as possible regardless of which market you're testing. A clearer graduation path from cub to partner — the existing stages work but the bear's communication of "why you advanced" can be much sharper. And the technical content explaining how the architecture works — including this page and the rest of the /learn series.
If you've used a retail AI trading tool and hit a version of the architecture story above, or you've raised a virtual pet and have opinions about what makes one stick, I want to hear it. Email hello@trader-bear.com or find me on X at @bearfounder.
30 seconds to name yours. Paper-money by default. The bear grows with you, across stocks, crypto, futures, and prediction markets.
Adopt a bear →