Why Claude Code Keeps Feeding You Bad Trades

Most bad trade ideas do not announce themselves as bad. They become convincing one small answer at a time.

That is what makes researching a hunch with Claude Code so appealing. Ask what supports it, and Claude can build the case with you. If the argument leans too heavily on the last few days of market prices, say so. Claude can revisit it, compare another way to make the trade, remember what bothered you, and explain what changed. Each answer can make the idea feel more considered. Add a broker, and the eventual order stays within reach of that same conversation.

Compared with hopping between a terminal, a notebook, a broker screen, and seventeen browser tabs, it is a relief. No pretending otherwise: we like working this way.

That same convenience is also how a weak assumption can acquire confidence, become a real trade, and lose money. If the data or calculation underneath the answer is wrong, the conversation does not have to feel wrong. It can sound thoughtful right until the order is sent.

Suppose Claude says implied volatility looks expensive. Fair enough. Expensive against what? Perhaps it fetched a recent history, cleaned a few gaps, chose a comparison window, ran a calculation, and interpreted the result. The answer arrives in a few seconds and flattens that whole chain into one confident paragraph. A stale close, an odd window, or two series measured at different times can quietly survive all the way into a very polished recommendation. This sounds pedantic right up until capital is attached to it.

You can ask Claude to show its work. You should. It may catch the problem and run the calculation again. Yet there is something slightly absurd about asking the same conversation to invent the research process, perform it, and then design the controls that tell us whether it can be trusted. We have turned one trade question into a small temporary market-data project. Tomorrow's question may produce a different one.

This is not a complaint about Claude. Reasoning through an idea in conversation is the useful part. Claude is good at following the thread when a trader says, “No, that is not what worries us,” or asks what evidence would kill the trade. It can surface assumptions that were previously just a feeling. It can make disagreement productive instead of sending us back to a blank screen.

A Claude subscription is still not a maintained market-research operation. Those are different purchases.

The broker does not close that gap either. A broker MCP principally gives the agent access to account operations and execution. It is the route to market, not a research department we get to presume exists. Once the analysis survives, making the order easier is excellent. It does nothing to repair the data or assumptions that produced the analysis.

The missing piece is the work most traders never want to rebuild in the middle of a conversation: keeping the data usable, defining calculations consistently, maintaining models, and producing outputs that mean the same thing when we return tomorrow. This is the layer Sharpe Two provides.

We describe it as deterministic market intelligence. The word can sound needlessly grand, but the idea is ordinary. The same defined inputs should run through the same maintained process and produce the same output. If a result looks strange, there is something concrete to inspect. We do not have to reconstruct a clever chat from memory and hope we ask the questions in precisely the same order.

Deterministic does not mean correct forever. Markets remain rude enough to embarrass good models, and a repeatable answer can still be wrong. What it gives Claude is a stable piece of work to question rather than a calculation it just made up for the occasion.

Forecasts are one example, when the trade actually needs a forward view. Claude can reason from a maintained Sharpe Two forecast, compare it with the trader's thesis, and explain the disagreement. It does not have to look at a recent chart and manufacture the missing bridge between “this happened” and “therefore this should happen next.” Probabilities can be useful in the same way. They are not the question every trader asks, and they are certainly not certainty wearing a decimal point. They are simply maintained inputs for the moments when odds matter.

This changes the conversation in a way we care about. Two trades with similar headline numbers may sit in very different conditions. Claude can explore why, pull on the part that looks inconsistent, and help the trader discover that the supposedly obvious setup is not obvious at all. The language model is doing what it does well. The market intelligence is not being recreated while it talks.

Sharpe Two intelligence is available through our API, and we are launching the Sharpe Two MCP because this maintained work should fit naturally inside an agent conversation. We want Claude to be able to reach for a defined market input as easily as it can reach for an account balance. That still does not make the whole system autonomous. It leaves more of the conversation for the questions that deserve judgment.

We stop before order routing. The trader still chooses the structure, considers the cost, sets the size, decides the allocation, and makes the rather consequential call to commit capital. The difference is what reaches that decision: Claude can push the idea further, revisit it from another angle, and keep the discussion anchored to maintained Sharpe Two intelligence instead of rebuilding the market premise as it talks.

This is where Sharpe Two earns its place in the conversation. Claude should not spend half the session rebuilding market intelligence and the other half defending what it just invented. Give it maintained research instead, and the conversation can finally move where the trader wanted it to go in the first place: from a question, to a view, to a trade worth putting on.