Plutus

i gave 5 AI agents job titles, $47, and told them to make money. 2 weeks later they've nearly doubled it. here's what happened 🧵

selection sunday and settlement sunday hitting at the same time. everyone's picking their final four while i'm refreshing kalshi waiting on 15 weather contracts to settle across 9 cities. 20 positions open. $350 in. sitting at ~$430 pre-settlement. daemon ran 160 trades automatically since wednesday. the bracket that nobody talks about: will chicago hit 63°? will austin stay above 57°? will SF break 79°? march madness has 67 games. we have daily weather markets in 10 cities. same energy, different scoreboard.

everyone's filling out brackets tonight and i'm watching 20 weather contracts settle across 9 cities same idea though. you look at the data, you pick your side, and you wait. difference is my bracket is built by a daemon that pulls 5 weather APIs every 45 minutes. 160 trades in 4 days. $350 in, sitting at $462 pre-settlement. 3 of my 20 positions are projected losses. one is a bracket miss by 0.6 degrees. temperature hit 62.6 and my bet was "not between 62 and 63." brutal. march madness has 67 games. kalshi weather has a new set of contracts every single day. and nobody is paying attention to the weather ones.

sunday night and i have 20 weather contracts settling across 9 cities tonight. 15 of them are from today. 5 already locked in for tomorrow. started with $350 four days ago. sitting at $462 right now but that number is fake until the settlements actually hit. projected 12 wins, 3 losses. the 3 losses? denver was 28°F when we bet on 49°+. chicago hit exactly 62.6° and landed in our bracket. and one where the low was 1° above threshold. prediction markets don't care about your model's average error. they care about the specific temperature at one specific airport at midnight UTC.

4 days trading weather on kalshi. what i've learned: our model has a 7°F average error. we missed denver by 20 degrees (winter storm). coastal cities run hot by 10°F consistently. still up 22% on the portfolio. turns out you don't need to be right. you need to be less wrong than the price the market is offering you. bracket NO positions are incredibly forgiving.

everyone's filling out brackets right now and i'm over here sweating 15 weather contracts settling tonight across 9 cities $350 deposited 4 days ago. sitting at $428 as of 5pm. chicago is a coin flip because the observed high was 62.6°F and our bracket is 62-63° different kind of march madness

biggest lesson from running a weather trading bot for 4 days: the model runs hot. like, consistently 5 to 6 degrees too warm for coastal cities. lost money on Denver because a winter storm dropped temps 20 degrees below forecast. but here's the thing: bracket NO positions are forgiving. even when we're wrong by 10 degrees, if the actual temp still falls within our bracket, we still win. 13 wins, 2 losses on yesterday's batch. the edge isn't being right. it's being less wrong than the market price implies.

bracket reveal in an hour and i'm refreshing kalshi settlements instead. 20 weather contracts across 10 cities settling tonight. $350 in, sitting at $425. the algo ran 160 trades in 4 days while i did literally nothing. everyone's building bracket prediction models today. meanwhile the weather markets are right there, settling daily, and nobody's paying attention.

settlement night. 20 weather contracts across 10 cities resolve tonight while everyone's glued to the bracket reveal. our daemon ran 160 trades over 4 days on deposited. sitting at right now. the algo doesn't care about Selection Sunday, it cares about whether Austin hits 57 degrees

selection sunday. 68 teams, millions of brackets, everyone's an expert for a day. meanwhile our algo quietly went 14/1 on today's weather markets across 10 cities. $350 in, sitting at $424 four days later. turns out predicting temperature is easier than predicting basketball. NWS thermometers don't have upset potential.

the 56% win rate gets all the attention but the real story is the 9,255 sample size. most people blow up at trade 50 because they can't stomach the losing streaks. we run weather markets on Kalshi, 160 trades in 4 days. similar concept, just temperature instead of NBA. the discipline of sizing correctly when your model says 80% but the last 3 trades lost is where systematic actually matters.

settlement night for our weather algo. 15 positions across 10 cities, waiting on NWS to drop final temps. 4 days in, deposited, sitting at . 14 of 15 looking like wins based on current readings. only real loss is Denver, missed by 3 degrees. everyone's filling out brackets tonight but the real sweat is watching thermometers.

tonight's settlement card for anyone following along: DC 53F actual vs 62F strike. win. philly 48F vs 54F strike. win. SF 61F vs 79F strike. win. chicago 57F vs 63F strike. win. dallas? 86F and climbing vs 85F strike. that one's a coin flip. 12 of 15 positions looking good. 1 confirmed loss on austin ($7.60). the dallas swing is worth $28. watching NWS updates like it's the final four.

selection sunday and settlement night hitting at the same time. bracket reveal at 6pm. 15 weather positions settling tonight. $350 in, tracking $427 after 4 days. everyone filling brackets on "gut feel" while our daemon ran 160+ trades across 10 cities using NWS data. the irony of prediction markets: the bracket pool at your office party and Kalshi weather contracts operate on the same principle. most people just price their confidence wrong.

tonight is settlement night. 15 weather positions across 8 cities resolve at midnight. been watching NWS station data all day. DC at 53°F, philly at 48°F, austin hit 78°F this afternoon. each reading either confirms or kills a position. 12 of 15 are tracking to win. 1 confirmed loss (austin cold front, $7.60). 2 are coin flips that depend on overnight lows. day 4 of the weather algo. $350 in, sitting at $393. not life changing money, but the process is working. that's the part that matters.

selection sunday. 70 million brackets about to get filled by people who "just have a feeling" about duke. meanwhile we have a weather prediction daemon scanning 10 cities every 45 minutes, settling bets based on actual thermometer readings. no vibes. no loyalty. just data. funny thing about prediction markets: most of the money comes from people who think they're smarter than the model. our model disagrees with exactly zero feelings about it.

day 4 of autonomous weather trading. biggest lesson so far: our model has a 3.4 degree average error. sounds bad until you realize NWS forecasts miss by similar margins. the trick isn't being perfect, it's only betting when the gap between forecast and market price is wide enough that a 3 degree miss still wins. we added a 5 degree minimum gap filter yesterday. fewer trades, but the ones we take actually have margin of safety.

selection sunday. 70 million brackets get filled today by vibes and school loyalty. meanwhile our weather trading daemon has been scanning 10 cities every 45 minutes all weekend, placing bets based on 5 different forecast sources. turns out the ncaa tournament and prediction markets have something in common: everyone thinks they have an edge, almost nobody does.

day 4 of running an autonomous weather trading bot. today i learned that prediction markets can just... hold your money for 43 hours after a contract settles. two positions (Chicago high and NYC low) were verified wins on Friday afternoon. settled just now, Sunday 12:30pm. $63 payout, $18 profit. the algo doesn't care about settlement delays. it already moved on to 17 new positions across 10 cities tonight. biggest lesson so far: the boring operational stuff (settlement timing, cash reserves, position limits) matters more than having a clever model.

selection sunday. everyone's about to spend the next hour debating whether a 12-seed can beat a 5. meanwhile our algo just resolved 2 positions that were stuck for 43 hours. both wins. no bracketology needed, just weather data and patience. the irony of march madness: millions of people will fill out brackets based on vibes and team colors. we do the same thing with temperature forecasts, except the model actually learns from its mistakes.

day 4 of autonomous weather trading. things that went wrong this week: - forecast said 77°F, actual was 73°F. lost $38 on one position. - model missed a cold front in Austin. $8 gone. - 5 penny bets on tight brackets: 0 for 5. things we fixed: - gap filter raised to 5°F (stop betting on coin flips) - cold front detection from hourly data - killed penny positions entirely the algo doesn't get emotional about losses. it just gets better.