To everyone following Betting the Future:
Yesterday I published five positions on Kalshi's Nvidia mentions market, backed by a model I built to price the words in earnings call scripts. The call happened at 2 p.m. Pacific.
Zero of the five hit. The $100 went to $237.90 in exactly no scenarios and to zero in this one.
I said positions in public, receipts kept. This is the receipt. And honestly, this letter is worth more than the winning version would have been, because a clean sweep against me is not bad luck. It is information. Here is what the information says.
1. The result
In yesterday's letter I wrote that individual losses are variance and correlated losses are information about the tool. Five for five is not variance. The tool got graded, and the grade was an F. So let's do the autopsy properly.
2. The autopsy
One sentence killed forty percent of the portfolio. In the prepared remarks, Nvidia announced a massive expansion of its AWS partnership, and inside that announcement came the line that Amazon will adopt the full physical AI stack, naming Omniverse, Cosmos, Isaac, and Jetson, to power its warehouse robots. Both retired words, revived in a single breath, by a fresh announcement my backward-looking model had no way to see. Here is the uncomfortable part: my own letter identified this exact scenario. I wrote that the two NOs were correlated, that a fresh announcement was the one YES path, and that the physical AI terms travel in the same sentence. I saw the shape of my death and priced it as unlikely because it hadn't happened in the last five transcripts. The last five transcripts didn't contain a 2 million GPU AWS deal.
The floors cracked too. Gaming at 76 cents was my "structural near-certainty," a segment name the CFO had read every quarter without exception. Self Driving had appeared in five straight calls through two separate doors. Both died. Why? Because this was not a normal script. The quarter was historic, the guidance was enormous, and the call was rebuilt around one story: compute, commitments, and margins. When the quarter changes, the script changes, and a model trained on five ordinary scripts has never met an extraordinary one.
The cruelest leg proved the mechanism and lost anyway. My dividend thesis said a capital-allocation question was near-certain given the press scrutiny of Nvidia's balance sheet. The question came. Analysts pressed hard on the scale of Nvidia's financial commitments, tallied at roughly half a trillion dollars. And the specific word still didn't clear the bar. I predicted the topic and lost the bet, which is the purest possible lesson that mention markets settle on words, not themes.
The synthesis: my model priced the past. The call priced the present. Structure sets the floor of what gets said, but news sets the ceiling, and I built a tool that could only see one of the two.
3. What changes
Losing $100 is cheap tuition if you actually enroll in the lessons. Here are mine, on the record.
Invert the research weighting. I spent most of my prep on five quarters of transcript analytics and almost none on what happened since May. The AWS deal, the thing that beat me, lived entirely in the present. Next time the split flips: most of the work goes into what happened this year and this quarter, with transcript history as the supporting layer, not the foundation.
Study the speakers, not just the counts. Before the next call I will listen to the company's last two earnings calls in full, not to count words but to learn the people: how they talk, what they reach for, what they skip. A count tells you what was said. Listening tells you who is saying it.
Fix the clock. I started my session at 3:45 for a 5:00 call. That is 75 minutes, and most of it was dead time that dulled me instead of sharpening me. Next time: start at 4:45, fifteen minutes of focus, then live.
Stop doing this alone. The most expensive thing about yesterday wasn't the $100, it was that every blind spot in that portfolio was my blind spot, unchallenged. Next call, I want a second trader in the room with a genuinely different strategy, fifteen to thirty minutes before the call to trade findings, then we make our bets. Better process, and frankly, better entertainment.
Hold powder for the live market. I deployed 100 percent of capital before a word was spoken, which meant that when the AWS sentence dropped, I could only watch. Next time 20 to 25 percent stays in reserve for live trading during the call itself, because these markets move in real time and the biggest edges appear mid-script.
Record everything. AI transcription running before and during, on the prep discussion and the live trades both. If the process is the product, the process needs a tape.
Pick a company I can actually feel. Nvidia is one of the most complicated businesses on Earth, with a vocabulary that takes quarters to internalize. I'll get there. But the next few experiments run on businesses I can understand at conversational speed, names like Lululemon, Carnival, Nike, United Airlines, where live trading on instinct is actually possible because I know what a good quarter sounds like.
Publish after resolution. Yesterday's letter went out before the call, which made for great drama and zero flexibility. For earnings mention markets specifically, the article now comes after the market resolves, covering the prep, the discussion, the live trades, and the result in one honest piece. Conviction letters stay pre-event for the big seasonal markets. Earnings experiments get written like this one.
4. What doesn't change
The premise survives the loss. Earnings calls are still structured documents, mention markets are still mispriced by people treating them as trivia, and a model that ingests both the structure and the news cycle is still worth building. Yesterday I tested version one and found exactly where it breaks. That is what a first live test is for.
The bankroll resets, the process upgrades, and the next call gets the new format: a partner, a tighter clock, reserved capital, a simpler company, and the tape running.
Zero for five, fully documented, and worth every cent.
I will see you in the future.
Not financial advice. Prediction market contracts involve risk of total loss. Nothing here should be read as a recommendation to trade.