Skip to content
Sign in
The Markets Desk — Week 44 · Fri 30 Oct 2026
Track record / Dated when made · scored when it resolves

The record of the judgment, not a quote screen.

Every view is stated in public with a date on it, and scored mechanically when it resolves — not before.

A machined steel tally counter with a copper thumb button, on paper.

Real return · the book

% only — sizing stays privateThe daily log

The headline here will be the return on the book, in percentages only, once the trade journal carries enough entries to plot. Until then the record below is the proof: open directional calls, each stating its direction and what would prove it wrong, scored when they resolve.

Call scorecard

Falsifiable directional calls, graded mechanicallyThe ledger
Total calls
13
Avg resolution
179d
Active now
12

Cap unset

The tripwire board

Every falsifier leg, by its last reading

Every leg of every falsifier in the Book on one board, sorted by its last reading: not what the prices did this week, but which of the reasons each call gave are still standing. Open any leg to read its call.

25 legs across 13 calls: 1 fired, 5 partly fired, 2 can't be checked, 1 not read yet, 1 can't happen now, 15 not fired.

Fired1

Partly fired5

Can't be checked2

Not read yet1

Can't happen now1

Not fired15

The clock and the grades

Each window, its reviews, and how a closed call is graded

Each call has a window, from the day it was published to the day it must be judged. The clock shows how much of each window has gone, every review as a tick, and a dot where a leg has fired. Beside it, how a closed call is graded: on the reason first, then the price.

KV cache before HBM46 Sept7 SeptLeg 1 · partly fired · 6 SeptLITE29 Apr3 Aug21 Aug29 Aug6 Sept9 OctLeg 1 · fired · 9 OctLeg 2 · partly fired · 29 AugRIGHT PRICE, WRONG REASONMiners turn to AI29 Apr6 Sept7 SeptPast parameter scaling29 Apr6 Sept7 SeptNVDA3 Aug6 Sept9 OctMU3 Aug6 Sept9 Oct29 OctSK HYNIX3 Aug6 Sept7 Sept9 Oct29 OctAVGO3 Aug6 Sept9 OctHybrid architectures6 Sept7 SeptNuclear scales6 Sept7 Sept11 SeptLeg 1 · partly fired · 11 SeptEnergy is the moat29 Apr6 Sept7 SeptInference costs −90%6 Sept7 SeptLeg 1 · partly fired · 6 SeptLeg 2 · partly fired · 6 SeptTSM3 Aug6 Sept23 Sept9 Oct29 Oct
Grades · reason against price
Right reason,price not yetnone yetThe goalnone yetPlain wrongnone yetRight price,wrong reasonLITEprice →reason →
13calls published
1closed, 1 as a thesis break
0graded right on reason
3open windows close on 13 Apr 2027

Theme calls have no price, so they are placed by their reason alone, on the reason edge at the left: the top half if the reason held, the bottom half if not.

Active calls

Stated at entry · next catalyst shown
Active calls, with the level each was called at
SymbolCallEntryNext catalyst
AVGOBroadcom: the right custom-silicon thesis at the wrong price392.16in 41d
000660.KSSK Hynix: the largest HBM rent, valued like a Korean cyclical2,215,000 KRWin 87d
MUMicron: priced like a cyclical at the top, contracted like a utility935.89in 55d
NVDANVIDIA: paying a market multiple for a franchise financing its own demand208.19in 19d
—Inference Costs Will Fall 90% by 2028 — The Memory Stack Is the Mechanism——
—Nuclear Is the Only Power Source That Scales for AI Data Centers——
—Transformer Successors: State-Space Models and Linear Attention Are Worth Watching——
—KV Cache Compression Will Cut Inference Costs 50% Before HBM4 Ships at Scale——
TSMTSMC: own the bottleneck, not the brand427.92in 76d
—Parameter Scaling Is Dead. Multi-Dimensional Scaling Is the New Paradigm.——
—BTC Miners Pivoting to AI Infrastructure: The Power Arbitrage——
—Energy Ownership Is the New AI Moat——

Where the record is strong

Resolved calls only

Accuracy by conviction

MEDIUM0% (0/1)

Accuracy by sector

AI Infrastructure — Optical Components0% (0/1)

Conviction-weighted index

Is high conviction actually working?
High Convictionn = 1
-2.2%
best -2.2%worst -2.2%
Medium Convictionn = 4
-16.3%
best +1.3%worst -42.4%
Low Convictionn = 0
—

Calibration

Probabilistic forecasts from the KB, Brier-scoredHow this works
Forecasts
5

Resolved

Most recent grades
Resolved calls
StanceTickerCallGrade
LONGLITELumentum: The Optics Play That Bridges Power and MemoryCalled Apr 13 at 871.18 · resolved Oct 9 at 1,085.42✗ Invalidated +24.6% in 179d
RESOLVED 2026-10-09 (our ruling). Tape: 871.18 → 1,085.42 (+24.6%). The falsifier's first limb fired: on the 11 Aug guide, co-packaged optics at scale moved to 2028, outside this call's window (closes 2027-04-13). The page said a one-year call cannot survive that, and it does not. The stock rose on lasers and transceivers (FY26 revenue +83%), not on CPO. Right price, wrong reason: recorded as a thesis break, not a win.
1 resolved callsMethodology

Review desk

9 of 12 due · reviewed weekly, on every catalyst, re-underwritten quarterly
Stale · 54dKV Cache Compression Will Cut Inference Costs 50% Before HBM4 Ships at Scalereviewed 54d agoLONG
Stale · 54dTransformer Successors: State-Space Models and Linear Attention Are Worth Watchingreviewed 54d agoWATCH
Stale · 54dEnergy Ownership Is the New AI Moatreviewed 54d agoLONG
Stale · 54dBTC Miners Pivoting to AI Infrastructure: The Power Arbitragereviewed 54d agoLONG
Stale · 54dInference Costs Will Fall 90% by 2028 — The Memory Stack Is the Mechanismreviewed 54d agoLONG
Stale · 54dParameter Scaling Is Dead. Multi-Dimensional Scaling Is the New Paradigm.reviewed 54d agoLONG
Stale · 49dNuclear Is the Only Power Source That Scales for AI Data Centersreviewed 49d agoLONG
Stale · 21dNVDANVIDIA: paying a market multiple for a franchise financing its own demandreviewed 21d agoLONG
Stale · 21dAVGOBroadcom: the right custom-silicon thesis at the wrong pricereviewed 21d agoWATCH

How the record is kept

Returns come from the trade journal, in percentages only. Every directional call locks at publication and is scored mechanically when it resolves — the stated direction against the realized move. A live call shows no running percentage on purpose: the position is exited on a broken thesis, not on price, so a number that moves daily would grade the wrong thing. Nothing is deleted; the resolved feed is the full history.