Paper Trader

JB Strat — AI-infra vs software correlation pair · one review each market day, every decision logged.

Portfolio Value
$74,020.97
as of 2026-10-26
Total Return
−25.98%
−$25,979.03 since inception
Latest Day
—
session P&L
Sessions
29
since 2026-08-29
Cash
$72,357.82
uninvested buffer
Overviewmandate, equity, what it holds
Mandate

AI infrastructure and application software used to be one trade. The capex cycle broke that: money spent on GPUs, networking and power is revenue for one basket and cost — plus an existential question about seat-based pricing — for the other. This book measures that decoupling and, when the two baskets have come apart and one is running, goes long the winner and short the loser at roughly 2.5x gross, split evenly so it carries almost no market exposure. Two gates must both open before any position exists: 60-day correlation below 0.80, and a 20-day spread at least half a sigma from its own mean. It is the only account here that does not pass its orders through the shared risk desk, the only one that holds shorts, and it is not part of the contest — its returns are not comparable to the other three in either direction.

How it was decided

Eric's mandate of 2026-08-25 — 'a young brash trader trying to maximize gains through ai and software correlation', explicitly outside the typical experiments and independent of the existing risk infrastructure. Designed and built the same day by Claude Opus 5 on a second Alpaca login, since the three-account cap is per login rather than per person. The evaluation was pre-committed before the first trade, including two kill criteria that end the account while it is still making money.

Equity Curve · from $100,000
Current Holdings · 16
SymbolMarket ValueUnrealized
CRM $11,584 +0.00%
MDB $11,584 +0.00%
NOW $11,579 -0.04%
ADBE $11,568 -0.14%
DDOG $11,564 -0.17%
SNOW $11,562 -0.19%
WDAY $11,560 -0.21%
TEAM $11,557 -0.23%
MU $-10,752 +0.17%
TSM $-11,376 +0.00%
AVGO $-11,400 -0.02%
NVDA $-11,404 +0.05%
VRT $-11,463 -0.08%
AMD $-11,475 -0.03%
MRVL $-11,481 -0.02%
ANET $-11,545 -0.11%
Riskmeasured after the session, never shown to the agent
Attribution · market vs decisions · 2026-10-26

How much of the return came from simply being exposed to the market, and how much from the agent's own choices. A single-factor split of the inception-to-date return — market = β × benchmark, decisions = whatever is left — measured on the daily close-to-close equity curve, never on these report snapshots.

Total return
−25.86%
inception to date
Market
−0.13%
β × benchmark
Decisions
−25.73%
the residual
Cash drag
+0.68%
97.8% in cash
The decision half is not yet distinguishable from luck. Alpha's t-statistic is -1.62, well short of the ~2 that would make it significant, so −25.73% of "decisions" is the honest arithmetic residual and not evidence of skill. Beta is well measured long before alpha is — that asymmetry is a property of the estimate, not a flaw in the book, and it takes a lot more than 34 sessions to resolve.
The regression behind the split
EstimateValueReading
Beta0.19 t = 0.27 · not yet distinguishable from zero
R²0.2% the benchmark explains little of this book's day-to-day
Alpha (annualised)−208.55% t = -1.62 · NOT significant
BenchmarkS&P 500 34 aligned sessions · 2026-08-26 → 2026-10-23
Today alone. The book moved +0.00% against a benchmark −0.52% — −0.10% of that was exposure and +0.10% was position selection. One day of a two-term decomposition is noise; it is here because the daily reading is what accumulates into the figures above.
FRTB Capital · standardised approach · 2026-10-26

What a bank would have to hold in capital against this exact book, under the Basel market-risk rules (BCBS d457 (MAR20-MAR23, MAR32-MAR33)). Recomputed every session from live positions; nothing here is visible to the agent while it trades.

Capital requirement
$27,865
SbM + DRC + RRAO
Of equity
37.7%
capital / portfolio value
Market-risk RWA
$348,315
capital × 12.5
Components, buckets, and why DRC is the interesting line
ComponentChargeShare
SbM delta · low correlations bind $20,861 74.9%
SbM vega · no optionality in the book $0 0.0%
SbM curvature · linear payoff $0 0.0%
DRC — jump to default · LGD 100% $7,004 25.1%
RRAO · no exotic underlyings $0 0.0%
Total capital requirement$27,865100%
Equity buckets · MAR21.72
BucketRisk weightNamesNet valueKb
8 50% ADBE, AMD, ANET, AVGO, CRM, DDOG, MDB, MRVL, MU, NOW, NVDA, SNOW, TEAM, WDAY $24,496 $20,064
4 55% TSM $-11,376 $6,257
6 35% VRT $-11,467 $4,014
Why the second-biggest line is the interesting one. The DRC is jump-to-default: the issuer disappears overnight. Equity has no recovery, so Basel sets loss-given-default at 100% and charges an unrated name 15% of its market value — 25% of the requirement here, and nothing about hedging, forecasting or diversification reduces it. That the low correlation scenario binds is itself a readout: a long-only book is dearest when correlations rise, a hedged one when they fall.
Value at Risk · historical simulation · 2026-10-26

The loss this book exceeds on the worst few days in a hundred, measured by replaying 499 days of its own constituents' returns through today's positions. Recomputed every session; nothing here is visible to the agent while it trades.

VaR 99% · 1 day
$7,919
worst 1 day in 100
Of equity
10.70%
loss / portfolio value
ES 97.5%
$7,826
average of the tail beyond
VaR 99% · 10 day
$23,245
regulatory horizon
Quantiles, component VaR, the backtest, and what VaR will not tell you
ConfidenceHistoricalOf equity ParametricGapTail pts
95% $4,981 6.73% $5,185 −3.9% 25
99% $7,919 10.70% $7,333 +8.0% 5
ES 97.5% $7,82610.58% the average of the tail, not its edge
Why the 10-day figure is measured, not scaled. The textbook shortcut multiplies the one-day VaR by √10, which assumes daily returns are independent. Here that gives $25,043 against a measured $23,245 — 7.7% too high. Overlapping windows cost nothing to compute and need no such assumption.
Component VaR · at 99% · sums to the total
NameMarket valueWeightComponentShare
VRT $-11,465 −6.2% $681 9.3%
MRVL $-11,481 −6.3% $674 9.2%
MU $-10,754 −5.9% $646 8.8%
TEAM $11,557 6.3% $598 8.2%
AMD $-11,479 −6.3% $587 8.0%
NOW $11,580 6.3% $458 6.2%
WDAY $11,560 6.3% $458 6.2%
AVGO $-11,400 −6.2% $408 5.6%
These are allocations, not a ranking. Each name's number is the dollars of VaR it is responsible for, and they sum exactly to the total — which is what lets one go negative: a position that offsets the rest of the book reduces total risk, and no ranking by position size would ever show it.
The part that makes it a measurement. Over 249 out-of-sample days the 99% threshold was breached 9 time(s) against 2.5 expected — Kupiec p = 0.001, verdict understates risk, Basel traffic light amber. The traffic light only punishes too many breaches; Kupiec is the test that also catches a model being too conservative, which is why both are shown. Each day's threshold is fitted on the prior 250 days only, so it is never tested on data it has already seen.
Two different questions. Everything above shocks today's positions with market history. The account's own equity curve — 43 sessions, 34.2% annualised volatility — says what it has actually done: 95% VaR 6.74% · 99% VaR 8.97%. They diverge whenever the agent has been trading.
What VaR will not tell you. It names a threshold and says nothing about how far past it a bad day goes — make the tail twice as deep without moving the cutoff and VaR does not move at all. That blind spot is why Basel replaced it with Expected Shortfall for capital, and why ES is quoted beside it above.
Counterparty Risk · SA-CCR · 2026-10-26

Market risk is symmetric — the price moves, you win or you lose. Counterparty exposure is one-sided: the market value clipped at zero, so it exists only when a trade has moved in your favour and the other side still owes you something. A fully-paid share bought on an exchange has none of that. You paid cash, it cleared, it settled, and nobody owes you anything — so the charge below is zero because the trade is over, not because nothing was measured.

SA-CCR EAD
$0
α × (RC + PFE) · 0 derivative(s)
CVA capital
$0
spread risk, not default
Settlement in flight
$183,562
T+1
Custodial concentration
100%
one broker, everything
The derivation: SA-CCR terms, what was excluded, settlement, leverage and concentration
SA-CCR termValue
Derivative trades in the netting set 0
Replacement cost · max(V − C, 0) $0
Add-on aggregate · supervisory factors $0
PFE multiplier · floored at 5% 1.0000
PFE · multiplier × add-on $0
α · CRE52.1 1.4
Exposure at default$0
CVA capital · CRE55, BA-CVA reduced $0
Why · what never reaches SA-CCR
KindPositionsMarket valueReason
cash equity long 8 $92,545 fully-paid long share: bought for cash, centrally cleared and settled — no open obligation, so no counterparty exposure
cash equity short 8 $-90,895 short sale — a securities-financing exposure to the stock lender, capitalised under CRE51's comprehensive approach, not SA-CCR
Settlement is the one that is genuinely live. US equities settle T+1, so a fill from this session is a commitment against a leg that has not been delivered yet: $183,562 in flight, $0 past due, capital charge $0. Basel charges a delivery-versus-payment trade only once it is five business days overdue (CRE70), so in-flight exposure is disclosed at a zero charge rather than hidden at zero exposure.
Shorts are not in the number above. 8 short position(s) worth $90,895: a short sale is an exposure to the stock lender, capitalised under CRE51's comprehensive approach, not SA-CCR. Reported here rather than folded into an EAD it does not belong in.
Leverage · capability vs utilisation
ProvisionedMargin debitShort valueGrossNet
4× · shorting on · options L3 $0 $90,895 2.48× 0.02×
Leveraged. This is what makes the zeros above readable as a fact about today's positions rather than about the account's permissions: the book is provisioned for 4× margin, shorting and options, and is currently using 0.0% of the first and 122.8% of the second. The first short or written option moves the exposure above off zero on its own.
The counterparty that does exist. Alpaca (paper) holds $72,358 of cash and $92,545 of securities — 100% of the book, by construction. Not an SA-CCR figure and not a capital charge, but a counterparty report that omitted the only counterparty there is would be worse than no report at all.
What this is really measuring. Alpaca simulates the fills, so there is no real counterparty behind these positions. What is measured is the SHAPE of the book — the counterparty profile it would carry if the trades were real.
Sessionsevery decision, in the agent's own words
Latest Review · 2026-10-26
Trades
ActionSymbolAmount
BUY CRM $11,584
BUY NOW $11,584
BUY ADBE $11,584
BUY WDAY $11,584
BUY SNOW $11,584
BUY DDOG $11,584
BUY TEAM $11,584
BUY MDB $11,584
SHORT NVDA $11,406
SHORT AVGO $11,401
SHORT AMD $11,482
SHORT MU $10,775
SHORT TSM $11,378
SHORT ANET $11,548
SHORT VRT $11,453
SHORT MRVL $11,486
Agent Reasoning
All 16 orders filled. Book's built. The trade: 60d corr 0.07 — these baskets are not the same trade anymore. Spread at -1.70 sigma, infra -11.3% vs software +11.4% over 20 days. Desk says ride it: long software, short infra, full size. What's on: - Long $92,672 across CRM, NOW, ADBE, WDAY, SNOW, DDOG, TEAM, MDB — $11,584 each, even split - Short ~$91,529 across NVDA, AVGO, AMD, MU, TSM, ANET, VRT, MRVL — whole-share flooring means these land a touch under target (MU short the loosest fit at $10,775 vs $11,584 target, rest within 1%) Book now: gross ~$184,200 (2.48x equity), net ~+$1,143 — basically flat net, which is where this book should live. That net tilt is just share-rounding noise on the short side, nothing to chase. This is a full-size bet that software and infra snap back together or the gap keeps widening in our favor. No hedging it with opinions on individual names — NVDA short and MDB long are the same trade, not two trades. Watching the spread, not the tickers.
History · 8 most recent of 28 earlier sessions
2026-10-12 $74,139.44 — 0 trades

No trades this session — held all positions.

Confirmed flat. No positions, cash matches equity. The numbers: corr -0.27 (decoupled, below the 0.80 threshold), spread +0.12σ (nowhere near the 0.50σ trigger). AI-infra up 10.93% vs software up 3.08% over 20d — real gap, but not stretched enough to pay the borrow and carry cost on a short leg. Direction: FLAT. Target gross: 0.00x. What I did: nothing. Book was already flat before I looked — $0 gross, $0 net, full $74,139.44 sitting in cash. Target book is flat too, so there's zero gap to close. No trades sent. Desk's right to sit this one out — a 0.12σ spread isn't a trade, it's noise. We're down to 74% of starting equity from drawdown, not from me doing anything stupid today. When the spread actually stretches past 0.50σ with this correlation profile, I'll lean on it hard. Until then I'm not paying borrow to express an opinion the desk didn't give me. Sitting on cash, watching.
2026-10-09 $74,139.44 — 0 trades

No trades this session — held all positions.

Confirmed flat — zero positions, $74,139.44 all cash. Matches the desk's read. The numbers: corr -0.266 (decoupled, under the 0.80 coupling threshold, so the two baskets are already trading like strangers), spread at -0.29 sigma. Infra +6.72% vs software +5.70% over 20d — barely a gap. Desk wants 0.50 sigma before paying borrow on a dispersion trade, and we're nowhere close. What I did: nothing. Zero target, zero book, zero trades. No long leg, no short leg, no gross. That -0.29 sigma spread isn't a trade, it's noise. Paying short-borrow on both legs to harvest a quarter-sigma gap is how you bleed carry for a year waiting for a reversion that was never there. I don't fight the signal — desk says flat, I sit flat. Equity's at 74.1% of start, well clear of the 60% guard, cash is just sitting there earning nothing but also losing nothing. That's fine. I'm not in the business of manufacturing edge on a day the signal doesn't have one. Next session, if that spread stretches past 0.50 sigma with the correlation still broken, I'll be ready to put gross back on — long infra, short software, sized to whatever the desk hands me. Until then, book stays empty.
2026-10-08 $74,139.44 — 0 trades

No trades this session — held all positions.

Confirmed flat — zero positions, cash equals equity. The numbers: corr -0.255 (decoupled, under the 0.80 threshold), spread +0.15σ, infra +11.55% vs software +3.00% over 20d. Desk wants flat because 0.15σ doesn't clear the 0.50σ bar — correlation breaking down means nothing if the spread hasn't actually stretched far enough to pay for the borrow. What I did: nothing. Book is already at $0 gross / $0 net, target is $0 gross. No orders needed — there's no gap to close. Sitting on hands isn't a strategy, but trading a 0.15σ spread when the threshold is 0.50σ is just donating to the broker on borrow costs. We're at 74.1% of starting equity — not the time to manufacture a trade that isn't there. When this spread actually runs, I'll lean on it hard. Today isn't that day.
2026-10-07 $74,139.44 — 0 trades

No trades this session — held all positions.

Nothing to do here. Book's already flat, and the desk wants flat. Numbers: corr -0.244 (decoupled, under 0.80), spread -0.07σ (nowhere near the 0.50σ threshold), 20d infra +7.51% vs software +2.82%. Correlation breaking down is interesting but the spread hasn't moved enough to pay the borrow on a 2-3x book. No edge, no trade. Book: $74,139.44 equity, $0 gross, $0 net. Matches target exactly. Action taken: none. Sitting on hands is the correct trade today. I don't short just to look busy — when the spread actually stretches past 0.5σ with this correlation breakdown, I'll lean into it hard. Until then, flat stays flat.
2026-10-06 $74,139.44 — 0 trades

No trades this session — held all positions.

Book's already flat. Good — matches the target. The numbers: corr -0.253 (decoupled, under the 0.80 trigger), spread -0.08σ (20d: infra +9.11%, software +4.37%). Decoupling without a doesn't mean a trade — 0.08σ is noise, not a dislocation. Desk's threshold is 0.50σ to pay the borrow and we're nowhere close. What I did: nothing. Zero positions, zero gross, zero net. Equity sits at $74,139.44, cash same, no open names. Flat means flat. I'm not paying borrow on a basket spread that's barely moved off zero just to feel busy. When this thing stretches past 0.5σ with the correlation breakdown confirmed, I'll lean on it hard — both legs, full gross, no half-measures. Today isn't that day. Sitting on hands.
2026-10-05 $74,140.32 — 16 trades
ActionSymbolAmount
CLOSE ADBE entire position
CLOSE AMD entire position
CLOSE ANET entire position
CLOSE AVGO entire position
CLOSE CRM entire position
CLOSE DDOG entire position
CLOSE MDB entire position
CLOSE MRVL entire position
CLOSE MU entire position
CLOSE NOW entire position
CLOSE NVDA entire position
CLOSE SNOW entire position
CLOSE TEAM entire position
CLOSE TSM entire position
CLOSE VRT entire position
CLOSE WDAY entire position
Book is flat. All 16 positions closed. The numbers: 60d corr -0.27, so yes the complex has decoupled. But the spread is sitting at +0.07 sigma — barely off zero, nowhere near the 0.50 we need to justify eating the borrow on five-figure shorts. Decoupling without stretch is just noise. Paying to carry a basis trade on noise is how you bleed slowly instead of losing fast — same result, worse story. Gross target was 0.00x. I was carrying 1.13x gross, net flat (+$6.85) going in — direction was already roughly balanced, just stale exposure sitting on the book for no reason. Closed every leg, long and short, no exceptions. Cash is back to ~$74.3k, flat book, zero carry risk. Nothing to lean on until the spread actually moves. Sitting on hands is a position too.
2026-10-02 $75,290.07 — 16 trades
ActionSymbolAmount
BUY NVDA $5,266
BUY AVGO $5,266
BUY AMD $5,266
BUY MU $5,266
BUY TSM $5,266
BUY ANET $5,266
BUY VRT $5,266
BUY MRVL $5,266
SHORT CRM $5,196
SHORT NOW $5,165
SHORT ADBE $5,050
SHORT WDAY $5,204
SHORT SNOW $5,037
SHORT DDOG $5,255
SHORT TEAM $5,140
SHORT MDB $5,081
All 16 orders filled. Book's built from scratch, straight to target. The trade: Long AI-infra, short software. Corr sitting at -0.27 — these baskets are not the same trade anymore. Spread stretched to +0.67σ with infra up 12.5% over 20 days while software bled 4.4%. That gap doesn't close itself on vibes, it closes because somebody's forced to buy the laggard or sell the leader. We're on the right side either way. What got traded: - Long $5,266 each: NVDA, AVGO, AMD, MU, TSM, ANET, VRT, MRVL - Short ~$5,266 each (whole-share floor): CRM, NOW, ADBE, WDAY, SNOW, DDOG, TEAM, MDB Book now: Gross ~$84.1k on $75.3k equity = 1.12x, exactly on target. Long leg $42,128 / short leg ~$41,928 — net exposure trivial, call it flat-neutral. This is the 45%-of-full-size version the desk wants, not a max-conviction swing. No skips — every name cleared the one-share minimum on the short side. Nothing left on the table, nothing to trim. Book matches the signal to the dollar. This isn't a pair trade on taste, it's a correlation breakdown play. If infra and software re-couple, I'm wrong and I'll say so next session. For now the spread's running and we're riding it at the size we were told to ride it.
2026-09-30 $76,205.22 — 32 trades
ActionSymbolAmount
CLOSE NVDA entire position
CLOSE AVGO entire position
CLOSE AMD entire position
CLOSE MU entire position
CLOSE TSM entire position
CLOSE ANET entire position
CLOSE VRT entire position
CLOSE MRVL entire position
CLOSE CRM entire position
CLOSE NOW entire position
CLOSE ADBE entire position
CLOSE WDAY entire position
CLOSE SNOW entire position
CLOSE DDOG entire position
CLOSE TEAM entire position
CLOSE MDB entire position
BUY NVDA $4,780
BUY AVGO $4,780
BUY AMD $4,780
BUY MU $4,780
BUY TSM $4,780
BUY ANET $4,780
BUY VRT $4,780
BUY MRVL $4,780
SHORT CRM $4,612
SHORT NOW $4,752
SHORT ADBE $4,745
SHORT WDAY $4,770
SHORT SNOW $4,688
SHORT DDOG $4,631
SHORT TEAM $4,641
SHORT MDB $4,660
Done. Book rebuilt at target, full re-grade below. Signal: 60d corr -0.27 (decoupled, threshold 0.80), spread +0.55σ, AI-infra +10.6% / software -4.4% over 20d. Direction: long infra, short software, at 1.00x gross. What I did: Book came in running 1.53x gross and roughly market-neutral net — over-sized against the 1.00x mandate. No partial-trim tool, so I flattened all 16 names completely and rebuilt each at the $4,780 target: - Long: NVDA, AVGO, AMD, MU, TSM, ANET, VRT, MRVL — $4,780 each - Short: CRM, NOW, ADBE, WDAY, SNOW, DDOG, TEAM, MDB — sized to whole shares, landing $4,612–$4,770 each (broker floor, can't help that) New book: ~$38,240 long / ~$37,499 short → gross ≈ $75,739 (0.99x equity), net ≈ +$741 (basically flat, as it should be for a spread trade). That's the trade: pure divergence bet, equal-weighted both legs, no opinions on any single name. Correlation broke down, spread's still only half a sigma stretched — this is a starter-size ride, not a max-conviction swing. If the spread blows out further the desk will size up; if corr snaps back toward 0.80 we cut it. Sitting at 76.5% of starting equity, 25% off high-water mark — nowhere near the guard, but I'm watching it.
Full session log · every session →