A FIVE-MINUTE BRIEFING ON THE MARKET CLOSE • REAL DATA ONLY • MATH DOES MATH

3PM Daily — The Market Close, in Five Minutes — From 800+ Stocks to 4 Setups

The market closes at 1:00 PM PT, but the work doesn’t stop. For the next two hours the big desks settle the day’s closing orders, earnings come out, and traders plan tomorrow. Most people who trade on their own miss all of it, or read about it hours later. 3PM Daily does that work for them: it scans 800+ stocks and publishes 4 setups by about 3:05 PM PT. The heavy math runs on my own machine; the briefing is served from the cloud (Cloud Run + Redis).

01 / DISCOVERY — WHY THE CLOSE IS HARD

The 3PM Gap: Wall Street's Afternoon Settlement Window

Wall Street does not shut down at 1:00 PM PT (4:00 PM ET). That is when the big desks process the day’s closing orders (Market-On-Close), earnings come out after hours, and tomorrow’s moves get planned. People who trade on their own log off at the bell, or get buried in long articles that arrive hours later.

3PM Daily scans 800+ stocks in that window and publishes 4 setups by about 3:05 PM PT. Getting ready for the close used to take about 45 minutes; the briefing takes under 5 to read.

Two Rules I Don’t Break

“A) REAL DATA ONLY: If a price appears, it came from actual exchange data, delayed at least 15 minutes. No placeholder prices, and no numbers made up by AI. B) MATH DOES MATH: A math library (TA-Lib) works out stops and targets. The AI only explains and debates — it never makes up a price.”

Real Data OnlyMath Does MathScore +4 or Higher Publishes
The Two Hours After the Bell — 1:00 PM → 3:05 PM PT
Runs Every Trading Day
✦
1:00 PM PT
1. Pull In the Close
~800 15-minute price bars, closing-auction prints, and ~4,000 insider filings (Form 4) go into the database.
◑
2:00 PM PT
2. Run the Math
TA-Lib scans ~827 stocks in 4.2 seconds and works out stops and support/resistance levels.
✧
3:05 PM PT
3. Publish the 4 Setups
The 4 setups that score +4 or higher are checked in the admin screen and published.
Score +4 or higher publishes • No badge is better than a fake badge
Problem: One dashboard for everything overloaded people — 5 different jobs were fighting for the same screen.
Solution: 5 separate screens, one per job, built only on real exchange data.
5 Jobs, Each With a Time Budget
PHASE 02 OF 06
Job // Time Budget
What They Need & The Screen I Built for It
1. Briefing Reader
5 min • Fatigued post-work
Wants 3 takeaways on why the market moved, without wading through tables. → A short dispatch + a market ribbon.
2. Tactical Hunter
15 min • Auction settlement
Wants entry prices, stops worked out from volatility (1x ATR), targets, and a way to test the trade on paper. → The setup list (4 setups, expandable to 98).
3. Rotator
10 min • Visual pattern seeker
Watches money move between the 11 S&P sectors to spot shifts early. → A market map + a sector strength table.
4. Catalyst Analyst
20 min • After-hours prep
Gets ready for tonight’s earnings and wants to hear both sides of an argument before deciding. → An earnings calendar + the AI debate panel.
5. Operator (Me)
30 min • Production publisher
Publishes the day’s briefing in under 30 minutes a day. The math runs on my machine; the site is served from the cloud. → The admin console.
*Disclosure: Names are representative composites to illustrate user jobs, not research participants.
Prices Pulled In: ~800 Bars
Filings Pulled In: ~4k Form 4s
Full Scan (4.2s): ~827 Stocks
Market Map: ~240 Tiles
02 / USER RESEARCH — 5 JOBS, 5 TIME BUDGETS

5 Jobs, Not 5 Personas

Designing one dashboard for a generic “retail trader” overloaded everyone. In 11 interviews, plus usage data from the first version, I found the real split is not who traders are — it is how much time they have at the close, and how tired they are:

Method
Workflow analysis + 11 trader sessions + usage data from the first version, across 5 different time budgets.
Key Insight
Made-up prices kill trust: Traders spot placeholder numbers right away and stop believing everything else on the page. If one price is invented, credibility drops to zero. Real exchange data only.
Decision
Built 5 separate screens, one per job, connected straight to the production database — no placeholder prices anywhere.
The Simple Version

Whether it is market data, budgets, or anything else: people do their best work when a tool shows them real numbers instead of guesses.

03 / SYSTEMS — MATH DOES THE MATH, AI DOES THE EXPLAINING

Math Does Math, AI Does the Explaining: How It Works in 4 Steps

One rule keeps the prices honest: math does the math, and the AI does the explaining. TA-Lib, a standard math library, calculates support, resistance, and stops straight from the data. The AI never touches those numbers — it only writes the context around them and debates what they mean.

What is TA-Lib?

TA-Lib (Technical Analysis Library) is a 30-year-old standard library for market math. It does not guess: if the numbers do not meet the criteria, it returns nothing, and nothing is shown. No badge is better than a fake badge.

How the Briefing Gets Made
TA-Lib scan in 4.2s • LangGraph
STEP 1
Pull In the Close
price_history_15m
~800 15-minute price bars, closing-auction prints, and ~4,000 insider filings go into the database (MySQL).
→
STEPS 2 & 3
Do the Math, Then Filter
TA-Lib + LangGraph
TA-Lib scans ~827 stocks in 4.2 seconds. Then 9 AI agents debate the risk.
→
STEP 4
Publish at 3PM
Cloud Run + Redis
Checked in the admin screen, then posted to the web, Discord, and X.
Every Table Has a ScreenNo table without a screen, no screen without a job
for_distribution
Setup List
Entry, stop, and targets (Hunter)
blog_content
The Briefing
The day’s story, 3 main takeaways (Reader)
industry_leader_reports
Market Map & Sector Table
Sector tiles and where money is moving (Rotator)
ai_debate_ledger
Debate Panel
The 9-agent debate and its verdict (Analyst)
Guardrail:If the math returns nothing, the badge is hidden. No badge is better than a fake badge.
the3pmdaily.com • One Layout, Five Screens
5 Screens // Nothing Extra
[MACRO FIRST]
ACTIVE
[TACTICAL CONSOLE]
[SECTOR BREADTH]
[CATALYST/WAR ROOM]
[NEWSROOM/ADMIN]
Detected Levels
Score ≥ +4 ships1x ATR Risk BoundMOC Settlement
The Day’s Setups
4 Setups • Stops Worked Out by Formula (1x ATR)
READY IN UNDER 5 MIN
!
Level Flag — Backed by TA-Lib Math
Several indicators agree (confluence score +4 or higher). The stop is worked out with a standard formula. Prices are never made up.
Stack
Next.js 15, TA-Lib, Cloud Run
Sources
price_history_15m + for_distribution
04 / IMPLEMENTATION — 5 SCREENS, ONE PER JOB

5 Screens, Each Built for One Time Budget

One Screen per Job
Each screen shows only what that job needs, for the time that person actually has:
01. Macro First (5 min)Short dispatch + market map (Reader)
02. Tactical Console (15 min)Setup list + $100k paper account (Hunter)
03. Sector Breadth (10 min)Market map + sector strength (Rotator)
04. Catalyst / War Room (20 min)Earnings calendar + AI debate (Analyst)
05. Newsroom / Admin (30 min)Review and publish (Operator)
// One layout for every screen
Every screen uses the same 12-column layout with fixed-height areas, so nothing jumps around when new data loads.
05 / AI — 9 AGENTS, EACH WITH A STRICT JOB

9 AI Agents, Each Arguing a Different Investing Style

Instead of one AI prompt cheerleading every stock, a LangGraph workflow runs 9 agents, each arguing from a different well-known investing style: buffett, graham, lynch, soros, burry, wood, dalio, oneil, and simons.

1
Value and Safety
Buffett + Graham look for strong profits (ROIC above 15%) and a margin of safety in the price.
2
Growth and the Big Picture
Lynch + Soros weigh growth against price (PEG under 1.2) and what the wider market is doing.
3
Debt and Downturns
Burry + Dalio probe debt, inventory build-ups, and where the credit cycle is heading.
4
Growth and Momentum
Wood + O’Neil + Simons test growth curves, breakout volume, and prices stretched too far in either direction.
Rule

Math works out the prices; the AI only explains and debates risk. If a number is missing, the badge is hidden. No badge is better than a fake badge.

Automated Tests That Run on Every Change (test_regression.py)
101/101 CHECKS PASS • 170ms
What runs: TA-Lib math • 50 test cases • 101 checks • the full run takes 170ms
The Guardrail, in CodeMath calculates, AI explains
1. The Math Runs First
taLib.findConfluence()
Worked out by TA-Lib in 4.2s
2. Nothing Passes, Nothing Shows
Score < +4 → null
No badge is better than a fake badge
3. AI Stays in Its Lane
9-Agent Debate
The AI only debates math that already checked out
✦See the Guardrail Code (level_guard.ts)▼
// Math works out stops and targets — the same way, every time
const levels = taLib.findConfluence(verifiedTape);
if (!levels) return null; // No badge is better than a fake badge

// LangGraph workflow: 9 agents debate the context
const debateLedger = await langGraphDAG.invoke({
  symbols: levels.passedHurdle, // Score >= +4
  agents: ['buffett', 'graham', 'lynch', 'soros', 'burry', 'wood', 'dalio', 'oneil', 'simons'],
  tape: verifiedTape
});

// The badge only shows when real math backs it
<LevelBadge visible={!!levels && levels.score >= 4} score={levels.score} />
Test run: 101 of 101 checks pass, in 170msSee the Full Test Setup in /lab →
Two Checks That Catch Bad Data
Type checks (50/50): Bad data is rejected at the door, before it can break the math. Missing-data check (50/50): If an indicator is missing, its badge is hidden instead of guessed.
The Speed Check
All 50 test cases and 101 checks finish in 170ms, so the tests run on every change. And if the math returns nothing, nothing gets published.
Outcome — Cost, and What Broke
Costs and Fixes, as They Happened
45 → <5 min
Time to Get Ready for the Close
Getting ready for the close used to take about 45 minutes of tab-hopping. The briefing takes under 5 minutes to read.
4 Setups
Out of 800+ Stocks Scanned
The scan takes 4.2 seconds: 800+ stocks in, the 4 setups that clear the bar out — each with entry, stop, and target.
~$18/mo
Cost to Run
The heavy math runs on my own machine, so the cloud bill stays at about $18 a month. It goes out every trading day.
What Broke — and How I Fixed It
  • Re-render loop: A data fetch inside a calendar widget kept re-rendering the page. Guarding the fetch with useEffect stopped the loop and the page stays responsive.
  • Unreadable tiles: Light-green tiles washed out the white text. Each tile’s brightness is now checked (getTileTextColor()) and the text switches to dark slate when a tile is too light to read.
  • A 600px empty gap: The two columns were unbalanced and left a big blank space on one screen size. Balancing the column heights closed the gap.
  • Old setups looked current: Debate results now load only when opened, and every setup is labelled Live, Settled, or Stale so its age is obvious.
What I Keep an Eye On
  • Usage signals: How often readers switch screens, open the full setup list, or start a paper trade — the signals that tell me the screens are doing their job. No personal tracking.
  • Serving: Briefings are served from a cache (Cloud Run + Redis), so pages open without re-running the math.
  • Publishing: One click in the admin screen posts the briefing to the web, Discord, and X.
06 / OUTCOME — WHAT IT COSTS AND WHAT I LEARNED

One Person, Run Every Day

I built 3PM Daily for my own close, and I run it every trading day. It reads real exchange data, does the heavy math on my own machine, and serves the briefing from the cloud for about $18 a month. The outcome I care about is the one at the top of this page: a close that used to take 45 minutes of tab-hopping now takes under 5.

Design
8 building blocks combined into 5 screens — one for each job, matched to how much time that person has (5 to 30 minutes). Nothing on a screen is there for decoration.
How It Runs
~800 15-minute price bars and ~4,000 insider filings come in; TA-Lib scans ~827 stocks in 4.2 seconds. Real exchange data only.
AI & Testing
A LangGraph workflow runs the 9 agents. The math is checked by 101 automated tests over 50 test cases, and the whole run takes 170ms — fast enough to run on every change.