DESIGN REFERENCE · LOCKED 23 AUG 2026 · THE BUILD MUST MATCH THIS PAGE · LOCKED CLICKABLE VISUAL PROTOTYPE — ILLUSTRATIVE, NON-LIVE RECORDS ONLY; NOT AN OPERATIONAL DATA SURFACE. OPEN THE REAL LAB ROUTE MAP & DOCS
trd.fun / lab
Paper trading lab
The only way is up.
Backtesting and strategy research on real market data. Sign in to open the lab.
Continue with Google
Same account as trd.fun — one sign-in for both. Signed in already? You go straight to the dashboard.
ACCESS IS APPROVAL-GATED · EVERY NUMBER FROM A LOCKED RECORD
lab.trd.fun · click the Google button to enter the prototype
trd.fun / lab
Dashboard
Strategies
Tournaments
Forward
Docs
sam@fnatic.com
Paper trading lab · lab.trd.fun
The Lab
Backtesting and strategy research on real SIP data — every number from a locked record.
DocsChangelogStrategies
+$11,532.57
best bot, exact-40 runA · rank 1 of 40BEST BOT EQUITY · 149D40 bots × 149 days = 5,960 bot-days tested
Five-check family tested: no edge — its gains concentrate in one session. First-hour continuation is locked and ready to test next.
— research tldr
Ideas tracked19
Screened4
Ready to test1
No edge5
Alpha ledger v4340/500bp
Forward tradingOff
Winning bots · exact-40 leaderboard
Click any bot to open its daily history and every trade.
full table (40) →
Grade
Bot
Stock
Equity · 149d
Avg daily %
Total P&L
Fills
A
Dip-touch buyer · busy tape · MU
stack4-touch-relvol125
MU
+0.07%
$11,532.57
46
A
Near-dip buyer · busy tape · MU
stack4-near-relvol125
MU
+0.06%
$10,256.19
114
A
Dip-touch buyer · normal tape · MU
stack4-touch-relvol100
MU
+0.06%
$9,601.79
48
A
Near-dip buyer · normal tape · MU
stack4-near-relvol100
MU
+0.06%
$9,032.09
120
B
Dip-touch buyer · busy tape · TSLA
stack4-touch-relvol125
TSLA
+0.04%
$6,950.86
38
Grades are relative rank inside this run · family verdict: no edge (one-day concentration) · new bots land here after every wave
Research cycles
Fri failedRepaired
Friday’s nightly cycle wedged at the close and was repaired the next morning. No session today (Saturday).
next cycle Mon 24 Aug · 22:30 UTC · includes Friday catch-upall cycles →
— nothing testing yet first-hour wave 1 starts when the evaluator build lands
Verdicts 6
Five-check stack
5,960 outcomes · 84.7% one session
no edge
Daily genome v1
13,407 attempts · 0 survivors
no edge
Trend / MA-cross
screen
Gap continuation
screen
SPY lead-lag
screen
Cross-sectional momentum
awaiting attestation · unpublished
parked
Your theses macro views — yours, not the lab’s; always benchmarked vs just holding
Memory upcycle · MU · SNDK · long · 12-mo
Hold the core, trade the swings on the way up.
overlay: not chosen yet · benchmark: hold the basket
Google model catalyst · GOOGL · long · event
Buy ahead of the next frontier-model release, trade the reaction.
overlay: candidate — first-hour continuation · vs holding
+ New thesis
Record a macro view — names, direction, horizon, why. The scout suggests overlays that fit.
Idea backlog recorded, not yet screened — promote one and it enters the board at quick check
Strategy scout · AI
Proposes new mechanisms nightly from the literature and market structure — proposals land here as ideas.
3 new this week
Post-earnings drift
Prices keep moving after the print. Who pays: investors who underreact to earnings news.
needs earnings calendar feed
next up ↑
Volume-shock reversal
Fade one-sided volume spikes. Who pays: panic sellers into the spike.
screen design not started
promote ↑
Opening range breakout
First 30-min range break with volume. Who pays: early faders caught wrong-side.
screen design not started
promote ↑
VWAP reversion
Fade stretched distance from VWAP. Who pays: chasers far from fair value.
also an overlay candidate for theses
promote ↑
End-of-day momentum
Last-hour continuation into the close. Who pays: traders squaring up before the bell.
screen design not started
promote ↑
Turn-of-month seasonality
Flows cluster around month boundaries. Who pays: calendar-driven rebalancers.
screen design not started
promote ↑
Vol-compression breakout
Squeeze, then trade the expansion direction. Who pays: sellers of quiet markets.
also an overlay candidate for theses
promote ↑
Pairs reversion
Co-moving pairs, trade the spread snapback. Who pays: one-legged momentum money.
screen design not started
promote ↑
Index-rebalance flow
Scheduled add/delete pressure. Who pays: forced index buyers and sellers.
screen design not started
promote ↑
Short-squeeze pressure
High short interest plus a catalyst. Who pays: trapped shorts covering.
needs short-interest feed
promote ↑
Lunchtime liquidity fade
Midday spread and depth behaviour. Who pays: impatient midday crossers.
screen design not started
promote ↑
Lifecycle: idea → promote → quick check → ready to test → testing → survivor / no edge. A dash means no data — never a zero.
trd.fun / lab
Dashboard
Strategies
Tournaments
Forward
Docs
sam@fnatic.com
Strategy register
Strategies
Every mechanism the lab has tried, is testing, or has parked — with the evidence. Click any line to drill into its runs, equity and day-by-day decisions.
BoardList
Tracked19
In play1
No edge5
Parked1
Backlog12
Your theses macro views — yours, not the lab’s; always benchmarked vs just holding
Memory upcycle · MU · SNDK
Early in a memory upcycle, 12-month horizon — big swings expected on the way up. Hold the core, trade the volatility.
overlay: not chosen yet benchmark: hold the basket
LongThesis
Google model catalyst · GOOGL
event-driven — buy ahead of the next frontier-model release, trade the reaction
LongThesis
+ New thesisrecord a macro view — names, direction, horizon, why. The scout suggests overlays that fit.
In play
First-hour continuation
Opening-hour drift persists into midday. Who pays: participants who fade the open too early.
quick check passed 100bp genesis · testing starts after build
Ask AIReady to test · locked 22 Aug
Tested — no edge
Five-check stack · 40 bots
RSI, Bollinger touch, relative volume, trend and session-window gates in confluence.
5,960 outcomes 84.7% of P&L from one session
No edge · 21 Aug
Daily genome v1
Overnight genome search over daily-bar rules. Nothing beat buy-and-hold on real data.
13,407 attempts 0 survivors vs buy-and-hold
No edge
Screened out at the mechanism gate ruled out before any ledger spend
Trend / MA-cross continuation
free-look event study · no edge after costs
No edge at screen
Gap continuation
free-look event study · effect not present
No edge at screen
SPY lead-lag
free-look event study · lag decays inside spread
No edge at screen
Parked
Cross-sectional momentum
A survivorship-free point-in-time corpus is sealed; the local result still awaits independent attestation and is not published.
survivorship-free corpus sealed · result awaiting attestation
Parked · evidence-unpublished
Idea backlog recorded, not yet screened — future strategies live here, not in docs
Strategy scout AI
Proposes new mechanisms nightly from the literature and market structure. Proposals land here as ideas with a quick check scheduled — chat with any strategy to explore it or spin up a variant.
3 new this week
Overnight drift
hold close → open · next in the screening queue
Idea
Post-earnings announcement drift
needs earnings calendar feed
Idea
Volume-shock reversal
fade one-sided volume spikes
Idea
Opening range breakout
first 30-min range break with volume
Idea
VWAP reversion
fade stretched distance from VWAP
Idea
End-of-day momentum
last-hour continuation into the close
Idea
Turn-of-month seasonality
flows around month boundaries
Idea
Vol-compression breakout
squeeze, then expansion direction
Idea
Pairs reversion
co-moving pairs, spread snapback
Idea
Index-rebalance flow
scheduled add/delete pressure
Idea
Short-squeeze pressure
high short interest plus a catalyst
Idea
Lunchtime liquidity fade
midday spread and depth behaviour
Idea
Lifecycle: idea → quick check → candidate → ready to test → testing → survivor / no edge. Parked = waiting on data. A dash means no data — never a zero.
trd.fun / lab
Dashboard
Strategies
Tournaments
Forward
Docs
sam@fnatic.com
Strategies/ Five-check stack
Tested family · canonical_149
Five-check stack
Enter only when five independent checks agree — RSI, a Bollinger-band touch, relative volume, a trend filter and a session window. Hypothesis: confluence gates filter noise into a small number of high-quality entries.
Who paysmomentum chasers arriving after the move is already confirmed.
Ask AIRules lockedNo edge · 21 Aug
No edge. 84.7% of the pooled family P&L came from a single session (20 Jul 2026). One lucky day, not a repeatable mechanism. No ledger spend beyond discovery.
— verdict, from the pre-registered (locked) test record
Bots40
Sessions149
Checked outcomes5,960
1-day concentration84.7%
Ledger spenddiscovery only
Family equity · pooled, all 40 bots
INDEXED TO 100 · JAN–AUG 2026
Bot drill · Dip-touch buyer · busy tape · MU
A · rank 1 of 40stack4-touch-relvol125HOVER A MARKER FOR THE TRADE · CLICK A DAY TO LOAD IT
FILLS 46P&L · LIFETIME+$11,532.57COSTS $2,304.80POSITIVE DAYS 16NO-TRADE DAYS 126
Win days16 of 23 traded
Gain/loss ratioavg win vs avg loss
Gains only
+$16,890
Losses only
−$5,357
Best day
+3.48%
Worst day
−1.90%
Max drawdown
−2.38%
In market
15% of days
Gross P&L · quarterly1.9k · 2.4k · 7.2k
Bot window · 13 Jul → 07 Aug
EQUITY · DAILY P&L BARS · THE STOCK UNDERNEATH — ONE SHARED TIMELINE
Session
Fills
Day P&L
Note
2026-07-16
0
0.00%
no entry — checks never aligned
2026-07-17
2
+0.42%
entered 10:12, target exit 13:05
2026-07-20
2
+3.48%
best day — entered 09:47, target exit 15:20, shown below
2026-07-21
2
−0.61%
entered 11:03, stopped out 14:40
2026-07-22
0
0.00%
no entry — session window closed first
2026-07-20 · session detail · MU
+3.48%1-MIN BARS · ALPACA SIP, LOCKED CORPUS
Decision trace
09:47:12 ET RSI(14) 28.4 < 30 — pass 1/5
09:47:12 ET lower Bollinger touched at 91.45 — pass 2/5
09:47:12 ET relative volume 1.31 ≥ 1.25 — pass 3/5
09:47:12 ET trend filter: inside 20-day VWAP band — pass 4/5
09:47:12 ET session window 09:45–15:30 open — pass 5/5
09:47:13 ET ALL CHECKS PASS — BUY 120 MU @ 91.42
15:20:41 ET target band reached — SELL 120 MU @ 94.60 · day +3.48%
Every number on this page is read from the locked run record. Nothing is recomputed at view time.
trd.fun / lab
Dashboard
Strategies
Tournaments
Forward
Docs
sam@fnatic.com
Research runs
Tournaments
Nightly research cycles on live sessions, and locked projections over historical corpora. Drill any run: run → bot → day.
Nightly cycles
22:30 UTC on trading days — a cycle that produced nothing says so.
MON 24 AUGScheduledRuns at 22:30 UTC · also catches up the repaired Fri 21 Aug session.
FRI 21 AUGFailedRepairedDispatch wedged at the close. Guard-verified repair applied 22 Aug; the cycle is created in Monday’s catch-up.incident →
THU 20 AUGCompletedCycle closed · no candidates promoted.open →
WED 19 AUGCompletedCycle closed · no candidates promoted.open →
Historical projections
Locked corpora, exact-minute completeness rule.
One row = one full campaign: a strategy family tested over a locked corpus. All 5,960 historical tests so far live inside the first row — open it for the per-bot leaderboard.
Run
Family
Scale
Outcomes
Best bot
Verdict
Date
canonical_149
exact-40
Five-check stack
149 × 40
5,960
Dip-touch buyer · MU +$11,532
No edge
20 Aug
open →
first-hour · wave 1
80 discovery attempts budgeted
First-hour continuation
—
—
—
Planned
pending build
Every future campaign — new waves, deeper history, thesis overlays — lands here as a new row.
Corpus library · 26 locked revisions · 2018–2026
Bad years included: 2018 Q4, the 2020 crash, the 2022 bear. Alpaca SIP source, locked under the exact-minute completeness rule.
browse corpora →
A run that failed says failed. Dashes mean no data — never zero.
trd.fun / lab
Dashboard
Strategies
Tournaments
Forward
Docs
sam@fnatic.com
Tournaments/ canonical_149 · exact-40
Historical projection · run leaderboard
canonical_149 · exact-40
149 sessions × 40 bots · 5,960 outcomes · graded on exact pooled return · 5,960 verified, 0 refused
LockedFamily verdict: no edge
Return
sum P&L / sum daily opening capital, truncated to 0.01% — not a compounded account gain.
Drawdown
daily-boundary evidence gate — not intraday, not live.
Fills
each modeled buy and sell counts once — not round-trips.
Grades
relative rank within this run: 1–4 A, 5–12 B, 13–28 C, 29–40 D. Not a profitability promise.
Grade
Bot
Stock
Avg daily %
Max DD
Total P&L
Costs
Pos / no-trade
Fills
Days
A
Dip-touch buyer · busy tape · MU
rank 1 of 40 · stack4-touch-relvol125
MU
+0.07%
2.38%
$11,532.57
$2,304.80
16 / 126
46
149
A
Near-dip buyer · busy tape · MU
rank 2 of 40 · stack4-near-relvol125
MU
+0.06%
7.28%
$10,256.19
$5,701.09
36 / 92
114
149
A
Dip-touch buyer · normal tape · MU
rank 3 of 40 · stack4-touch-relvol100
MU
+0.06%
2.38%
$9,601.79
$2,403.59
17 / 125
48
149
A
Near-dip buyer · normal tape · MU
rank 4 of 40 · stack4-near-relvol100
MU
+0.06%
6.88%
$9,032.09
$6,000.46
39 / 89
120
149
B
Dip-touch buyer · busy tape · TSLA
rank 5 of 40 · stack4-touch-relvol125
TSLA
+0.04%
1.15%
$6,950.86
$1,902.87
14 / 130
38
149
B
Strict near-dip buyer · normal tape · MU
rank 6 of 40 · tied · stack5-near-relvol100
MU
+0.04%
2.38%
$6,466.75
$1,102.64
8 / 138
22
149
B
Strict near-dip buyer · busy tape · MU
rank 6 of 40 · tied · stack5-near-relvol125
MU
+0.04%
2.38%
$6,466.75
$1,102.64
8 / 138
22
149
B
Dip-touch buyer · normal tape · TSLA
rank 8 of 40 · stack4-touch-relvol100
TSLA
+0.03%
2.13%
$5,654.51
$2,702.00
18 / 122
54
149
B
Dip-touch buyer · busy tape · NVDA
rank 9 of 40 · stack4-touch-relvol125
NVDA
+0.02%
1.97%
$3,678.68
$1,803.84
10 / 131
36
149
B
Strict near-dip buyer · busy tape · TSLA
rank 10 of 40 · stack5-near-relvol125
TSLA
+0.02%
0.56%
$3,423.07
$1,001.03
6 / 139
20
149
B
Strict near-dip buyer · busy tape · NVDA
rank 11 of 40 · stack5-near-relvol125
NVDA
+0.02%
0.61%
$3,175.51
$902.68
6 / 140
18
149
Drill path: run → bot → daily history → any day’s decision trace. Showing 11 of 40 — full table scrolls.A grades are relative rank inside this run — the family verdict is still no edge (one-day concentration).
trd.fun / lab
Dashboard
Strategies
Tournaments
Forward
Docs
sam@fnatic.com
Strategies/ Overnight drift
Idea · recorded 22 Aug — not yet screened
Overnight drift
Most of the market’s long-run return shows up between the close and the next open, not during the trading day. The idea: hold from close to open, stay flat intraday, and skip the intraday chop entirely.
Who paystraders who sell at the close to avoid overnight risk — they hand over the overnight premium.
Ask AIIdea
How it would trade
GREEN = HELD OVERNIGHT · DASHED = FLAT DURING THE DAY
The claim in one picture: the gaps between close and open carry the return; the trading day is mostly chop. This strategy owns the gaps and skips the chop.
When it wins · when it loses
Should win when
Steady uptrends with calm nights · broad indexes and mega-caps · news that lands after hours and gaps up.
Should lose when
Gap-down regimes — overnight shocks hit at full exposure · earnings nights on single names · crisis tape like March 2020, when the night IS the risk.
What we’re looking for
1Overnight share of returns stays positive across 2018–2026 — bad years included, not just the bull runs.
2The edge survives paying the spread at the open every single day — costs kill most versions of this idea.
3Not concentrated in a handful of lucky nights — we already refuted one one-day wonder.
What the quick check will test
• Split every corpus session into overnight vs intraday return, per symbol.
• Ask whether the overnight share survives real costs and spread at the open.
• Free look only — no trading simulated, no budget spent, no parameters tuned.
Status: queued — next in the screening queue · corpus: 26 locked revisions, 2018–2026
If it passes
It becomes a candidate: rules get written and locked before any test runs, it gets its own small test budget, and results land on this page — win or lose.
Lifecycle: idea → quick check → candidate → ready to test → testing → survivor / no edge
Every idea gets this same template: mechanism diagram · win/lose regimes · what we’re looking for · the quick-check design.
trd.fun / lab
Dashboard
Strategies
Tournaments
Forward
Docs
sam@fnatic.com
Strategies/ Memory upcycle
Thesis · yours, not the lab’s
Memory upcycle
MU and SNDK are early in a memory upcycle. Twelve-month horizon, high conviction on direction — and a lot of volatility expected on the way. The plan: hold the core, and trade the swings around it.
MUSNDKLong12-mo horizon
Ask AIThesis
The lab never grades your belief — it tests one thing: does trading around the position beat just holding it? If not, the honest answer is: hold.
— the benchmark that keeps this honest
How the lab trades a thesis · three layers
1 · Core
Hold the basket. This IS the thesis — it makes money if you’re right about direction, full stop.
2 · Overlay
A tested mechanism trades the swings around the core — add on dips, trim on rips, inside bands. Only mechanisms that passed the gates qualify.
3 · Benchmark
Always scored vs buy-and-hold of the same basket, after costs. The overlay must EARN its complexity.
Overlay candidates
VWAP reversion · add on stretched dips
Idea · needs screen
Vol-compression breakout · add on squeezes
Idea · needs screen
First-hour continuation · intraday timing
Ready to test
Planned test · regime-matched
Backtest the overlay on trending, high-volatility names from the locked corpus — NVDA 2023, TSLA 2020, AMD 2021 — and ask: when the direction call is right, does trading the path beat holding?
Status: queued behind the first-hour build · then forward paper-trades on MU + SNDK, scored vs hold
Thesis P&L will read: core + overlay vs pure hold — so you always see whether the trading added anything.
trd.fun / lab
Dashboard
Strategies
Tournaments
Forward
Docs
sam@fnatic.com
Live path
Forward
Forward paper-trading against the live market — the last gate before anything real.
Off
Research mode
The lab is in research mode. Nothing has earned forward trading yet — that is the system working, not failing. Every no-edge verdict here is a strategy that never got the chance to lose real money.
What turns it on
1A strategy survives its locked test — a real edge, not one lucky day.
2It posts at least 126 sessions of forward paper evidence without breaking its own rules.
3You approve it — promotion is always a human decision, never automatic.
Current candidates for this path: none. First-hour continuation would be next if it survives testing.
trd.fun / lab
Dashboard
Strategies
Tournaments
Forward
Docs
sam@fnatic.com
Reading room
Docs
How the whole system works, in plain English — dig in from here.
How it works · ELI5
The plan
Changelog
How the lab stays honest
Glossary
How the lab works
Robots practice trading with pretend money on real market data — and we keep score so honestly that a “no” means something.
1
An idea gets written down.You or the AI scout record a trading idea in plain English — plus who’s on the other side paying for it. Ideas cost nothing and sit in the backlog.
2
Quick check.A cheap look at history: does the claimed effect even exist? No trading simulated, no budget spent. Most ideas die here — that’s the point.
3
Rules get locked.Before any real test, the rules are written and frozen: which data, what counts as a win, how many tries allowed. Like a scientist pre-registering an experiment — nobody can move the goalposts after seeing results.
4
Tournaments.Many bot variants trade the idea over locked historical data — the last run was 40 bots × 149 days = 5,960 bot-days. Every simulated day is saved with every decision, so you can drill into any bot on any day and see exactly why it traded.
5
Verdict.Survivor or no edge — and the reason is written down. “No edge” is a win too: it’s a strategy that never got the chance to lose real money.
6
Budgets stop self-deception.Each family gets a fixed number of attempts. When it’s spent, no more tries — so we can’t torture the data until something “works”.
7
Everything vs just holding.The benchmark is always buy-and-hold. If a strategy — or a thesis overlay — doesn’t beat holding after costs, the honest answer is: just hold.
8
Forward, then (maybe) real.Survivors paper-trade the live market for 126+ sessions. Only then — and only with your approval — does anything touch real money. Nothing has earned it yet.
Ask the Lab◀ expand
Ask the Lab
FIVE-CHECK STACK · ANSWERS FROM THE LOCKED RECORD
×
run canonical_1495,960 outcomesday 2026-07-20
If the family has no edge, why did the top bot make $11.5k?
Rank is relative inside the run. stack4-touch-relvol125:MU earned most of its +$11,532 on 20 Jul — the same session that carries 84.7% of the whole family’s P&L. Remove that one day and the family is roughly flat after costs. One day isn’t a mechanism, so the verdict stays no edge.
leaderboard · rank 1day 2026-07-20
Worth testing a variant that only trades days like 20 Jul?
That’s a new idea — a regime filter: only trade when relative volume and gap size look like 20 Jul. I can record it as an Idea and schedule a quick check. It gets no test budget until the check passes.
Add to backlog as Idea
Show the worst dayCompare vs buy-and-holdWhy did 21 Jul stop out?
Ask anything about this strategy…
Answers cite the locked records they come from. The AI can propose ideas — it can never grant budget or change a verdict.