What’s new

The conviction auto-tuning that calibrates against your own track record now fits a separate pattern for each tier — direct prediction-market bets, indirect equities, and third-order plays — instead of pooling them into one curve. A conviction of 8 means different things across those tiers, and your sizing now reflects that (each tier that has enough resolved history gets its own correction; the rest fall back to your overall pattern). The "% sharper" figure on the conviction-check banner and /stats is now cross-validated — measured on picks held out of the fit rather than the same ones it learned from — so it tells you whether auto-tuning actually generalizes instead of flattering itself.

The matcher that links a direct pick to its Polymarket market got sharper. It now scores candidates in a length-normalized way, so a long, wordy, unrelated market can no longer outscore the right one just by containing more words; it penalizes wrong-year matches on recurring markets (so a 2026 question stops matching the 2024 version); it reads the event title, not just the market question, where the distinguishing detail often lives; and on a tie it prefers the more liquid market. The result is fewer "no price" legs and fewer wrong prices feeding your edge math.

Reworked how analyses and red-team challenges talk to the model. Output is now schema-constrained, so the occasional "model returned invalid JSON" failure is gone; transient overload errors retry automatically instead of killing your run (you'll see a brief "model is busy" instead of a hard error); and the large shared prompt is cached for faster, cheaper responses. No change to the picks themselves — this is purely about runs that complete cleanly.

ThesisTrader can now execute the equity picks it generates, not just suggest them. Connect Robinhood from the terminal header — a standard OAuth flow into Robinhood's agentic-trading account, with no passwords ever touching this app — and every indirect play and third-order equity card grows a Buy / Sell on Robinhood button. Click it, choose a dollar amount or a share count, confirm, and the order is placed through Robinhood's MCP trading server in the dedicated account you funded for the agent. Direction is read straight from the pick (Long → buy, Short → sell). Per-order and daily caps bound the blast radius, OAuth tokens are encrypted at rest, and disconnecting (one click in the header) revokes access immediately. Crypto and prediction-market legs stay informational — this is equities only.

Direct prediction-market bets now show three probabilities side-by-side: the market-implied price (e.g. 52¢), what the LLM thinks (your conviction), and the Bayesian posterior that fuses them. EDGE is the real number — how much your view actually differs from what's priced in, in percentage points. Equity and crypto legs show the same logic with assumed up/down move sizes per asset class. Every card also surfaces a fractional Kelly suggestion (raw Kelly × a fraction, capped per leg) you can snap to with a single APPLY button.

The portfolio simulator gained two principled ways to size your book. APPLY KELLY TO ALL snaps every slider to its fractional-Kelly suggestion (toggle ¼/½/full, cap per leg 10/20/50%). APPLY MV WEIGHTS uses Black–Litterman to overlay your conviction-driven views on a no-drift equilibrium prior, then reads off jointly-optimal mean-variance weights with the empirical covariance baked in — so two correlated long bets shrink, two hedged bets don't. A strategy chip in the simulator header (KELLY / BLACK–LITT / MANUAL / EQUAL SPLIT) shows what your current sliders represent, with a RESET button to return to equal split.

Below the expected/bull/bear bracket boxes is a new Monte Carlo panel: 5,000 correlated paths drawn from a 90-day return covariance, with horizon picker (1W / 1M / 3M). Reports P5 / P50 / P95, VaR-95, CVaR-95, and the probability the book loses money. The correlation matrix is now empirical too — cells based on real 90-day returns are underlined; the rest fall back to the old tier-based heuristic.

Once you've decided ≥15 picks (Took it / Passed in /history), the analyzer now recalibrates the LLM's conviction using your own results via isotonic regression. A new banner above the results shows your Brier score, the base-rate baseline, and Brier Skill Score with hover tooltips. An ISOTONIC ON/OFF toggle applies the recalibration to every card — you'll see "LLM 80% → 55%" when an over-confident call gets shrunk to the empirically-observed rate, with the Kelly suggestion shrinking accordingly. The /stats page gained a Probability calibration section with a reliability diagram bucket-by-bucket.

Ran the new math through correctness, regression, and UX audits and fixed everything they caught. Short bets in the Monte Carlo are now driven by your thesis EV instead of bleeding against historical drift. The Black–Litterman view value lives on the same log-return scale as the covariance (was off by orders of magnitude for crypto). Calibrated probabilities at the high end of a merged bucket no longer get silently clamped to the low end. Share links carry your Kelly fraction, max-position, horizon, τ, and δ settings instead of resetting recipients to defaults — and they no longer leak your personal calibration stats. APPLY MV WEIGHTS now zeros every non-BL leg so it can't silently over-leverage your book on top of existing direct bets. The sticky simulator is keyboard-navigable. Jargon labels (BSS, Π, Q, Ω, VaR, CVaR, τ, δ) gained plain-English hover tooltips. The two APPLY buttons no longer look identical.

Every analysis now extracts its catalysts — the dated events that confirm or kill the thesis, like an FOMC meeting, a CPI print, an election, or a scheduled earnings date. They show as a timeline on each thesis in /history, and a 📅 Catalyst soon tag flags any thesis with an event inside two weeks (filter the whole list by it with the Catalysts chip). On watched theses, a catalyst entering its final week becomes a watch alert — badge, and a line in the daily email — so the question stops being "did the price move" and becomes "when do I find out." FOMC, CPI, jobs-report, and company-earnings dates are pinned to the real published schedules — the Fed's calendar, the BLS schedule, and live earnings data — so they're exact, not guesses; only the unpredictable long tail like court dates and regulatory deadlines stays a model estimate, marked "est."

The daily watch digest now has an off switch. The History page shows a Daily email digest panel that spells out plainly what gets you onto the list — favoriting a thesis, or marking one of its trades Took it — and lets you turn the emails off (or back on) with one click. Every digest email also carries a one-click unsubscribe link in the footer. Turning emails off only mutes email; the ⚠ price-move alerts still show in-app on History.

Favorite a thesis, or mark one of its trades as Took it, and ThesisTrader now watches it. Once a day it re-checks the prices behind every pick and flags the ones that moved — a prediction market that swung 10 points, a stock or token up or down 10%, or a Polymarket question that resolved. Moved theses get a ⚠ badge in /history and a Watch alerts filter; expand one to see exactly what moved, from where to where, and whether it went your way. You also get a daily email digest of anything new, so a thesis you ran weeks ago can still reach you when the world shifts under it. Dismiss an alert and it stays quiet until prices move further.

New /stats page (linked from the terminal and history headers) rolls every tracked thesis into one calibration view. See your hit rate on the calls you actually took, broken down by tier — direct bets, indirect plays, third-order. A conviction calibration section checks whether the model's high-conviction picks actually beat its low-conviction ones, and tells you straight if they don't. Degen mode gets weighed against normal mode, and a "road not taken" section counts how many picks you passed on that would've won. This is the loop closing: per-thesis P&L showed you each call, /stats shows you the pattern.

Every pick in /history now shows its live P&L, computed from the prices captured the moment you ran the analysis. Tag each one Took it or Passed — won/lost is figured out automatically from the price move (Long / Short / YES / NO factored in), and resolved Polymarket markets are detected and marked Won or Lost terminally. Trades you passed on still get their verdict in muted styling, so you can see which calls you skipped that would've paid — that's the data that actually trains your judgment over time.

Every thesis you run is now saved to your account. Visit /history (link added next to ← EXIT in the terminal header) to scroll through everything you've analyzed, with the original prompt, date, and the full set of picks the model returned for each one.

On any past thesis you can star it to come back to later (filter via the Favorites chip), rerun it to load the original prompt and degen setting straight back into the terminal, or copy a share link that opens the full result view for anyone who clicks it.

Added a Buy Me a Coffee button at the bottom of the landing page and the terminal. If the tool saves you from one bad trade, buy me a cappuccino with the savings.

The help popover still claimed degen mode "favors volatile small-caps and memecoins" — but memecoins were dropped two days ago. Updated the tooltip to reflect the real behavior: leveraged ETFs, high-beta small-caps, and infrastructure tokens.

Added Google OAuth via Auth.js. The landing page stays public so you can read about the project, but running an analysis now requires signing in. After login you land straight in the terminal instead of bouncing through the landing page again.

Access is gated to an allowlisted set of emails as defense-in-depth. If your Google account isn't on the list, you'll see a rejection message with a link to an application form at /apply. Submissions show up in an /admin review page where approved applicants are added to the allowlist.

The server-rendered clock on the public landing page was flashing a hydration warning on load. Fixed.

Slimmed down the system prompt without changing the rules. Analyses return faster, and the occasional 504 timeout error on slow responses should be much rarer.

Removed memecoin suggestions from high-risk analysis mode. It still surfaces asymmetric, high-convexity bets — leveraged ETFs, high-beta small-caps, pre-revenue names, and infrastructure tokens with specific theses — without the reputational risk of PEPE and friends.

The site is now live at thesistrader.com woohoo! We're all grown up.

Added a short explanation above the correlation matrix clarifying that both axes show the same legs, and that cells show how the intersecting pair move together.

For theses where crypto doesn't genuinely fit (geopolitical conflict, recession, sector-specific plays), the analyzer now reaches for stocks, commodities, or FX instead. Each third-order pick also has to justify why its asset class belongs, so you stop seeing Solana recommended as a "Taiwan invasion hedge."

Fixed a pattern where the analyzer was defaulting to Bitcoin as a third-order pick on theses where it didn't belong (anything involving a stock-market bear case, for instance). The model now has to justify the asset class itself, not just the asset, and each third-order pick is labeled as AMPLIFIER, HEDGE, or CATALYST so its actual relationship to the thesis is visible.

Tapping the help icon next to the DEGEN toggle now shows the full explanation on any screen size. Previously the popover was clipped on narrow viewports.