A Discord bot that turns twenty independent traders into one daily signal —
wisdom-of-crowds, applied to the Nasdaq 100 futures open.
In 1906 Francis Galton watched 787 fairgoers guess the weight of an ox. No single guess was right. The median of all of them was off by one pound.
James Surowiecki's The Wisdom of Crowds names the four conditions that make this work: diversity of opinion, independence of judgement, decentralisation of knowledge, and a mechanism for aggregation. Break any one — especially independence — and the crowd stops being wise and starts being a herd.
Trading Discords break all four by design. Everyone reads the same chart, in the same channel, in real time, and the loudest voice anchors the room.
This bot rebuilds the four conditions mechanically:
| Condition | How the bot enforces it |
|---|---|
| Diversity | The roster is deliberately composed across eleven schools of market analysis — ICT, Wyckoff, Volume Profile, order flow, quant, macro, discretionary tape reading. Uncorrelated reasoning is the whole point. See docs/methodology.md. |
| Independence | The question arrives as a private DM. Nobody sees the running tally. There is no group chat before the cutoff. |
| Decentralisation | Every member trades their own book with their own framework; nothing is coordinated. |
| Aggregation | One scheduled job counts the votes, weighs the confidence, classifies the signal, and posts a single embed. |
No individual vote is ever shown. Votes are immutable once cast. Late submissions are impossible. Those aren't UX choices — they're the experiment's controls.
sequenceDiagram
autonumber
participant S as Scheduler (ET)
participant B as Bot
participant M as Members (DM)
participant C as Verdict channel
participant A as Admin
Note over S: 09:20 ET — trading days only
S->>B: job_broadcast_question
B->>M: DM: one question, four buttons
M-->>B: prediction, then confidence 1-10
Note over B: UNIQUE(date, member) — one vote, no edits
Note over S: 09:28 ET — cutoff, submissions close
Note over S: 09:29 ET — one minute before the bell
S->>B: job_post_consensus
B->>B: aggregate, classify, persist
B->>C: consensus embed
Note over S: 11:00 ET — session over
S->>A: DM: record the actual outcome
A->>B: /add_result, into the accuracy ledger
The eight-minute window (09:20–09:28 ET) is deliberate. Long enough to form a read; too short to go and ask somebody else what they think.
🟢 HIGH CONVICTION — EXPANSION UP
📈 Expansion Up 📉 Expansion Down ↔️ Expansion Both13/18 (72.2%) 2/18 (11.1%) 2/18 (11.1%)
⬜ Range 🎯 Avg Confidence 👥 Submissions 1/18 (5.6%) 7.4/10 18/20
Green for high conviction, gold for moderate, red for stay out. Red is a valid answer — a crowd that hasn't converged is information, not failure.
Evaluated top to bottom; first match wins (src/domain/classifier.py):
| # | Condition | Signal |
|---|---|---|
| 1 | Turnout below 50% of the active roster | 🔴 NO_SIGNAL |
| 2 | Average confidence below 5/10 | 🔴 NO_SIGNAL |
| 3 | Leading option ≥ 60% and average confidence ≥ 7 | 🟢 HIGH |
| 4 | Leading option ≥ 50% | 🟡 MODERATE |
| 5 | Anything else | 🔴 NO_SIGNAL |
Quorum is a proportion of the live roster rather than a hard-coded number, so the gate stays honest as the council grows or shrinks. Accuracy statistics count only HIGH and MODERATE days — scoring the days the bot explicitly refused to call would flatter the numbers.
- Scheduled DM broadcast to every active member on weekdays, with an NYSE holiday calendar and an admin
/skip_todayoverride. - Two-step button submission — direction, then confidence 1–10. Built on Discord persistent views, so a restart mid-window doesn't leave dead buttons in anyone's DMs.
- Immutable votes, enforced at the database level rather than only in the UI.
- Idempotent jobs. A delivery log prevents double-DMs; a
posted_to_groupflag prevents double-posts. Restart the process as often as you like. - Timezone-explicit scheduling. Every cron trigger is pinned to
America/New_York, so the bot behaves identically on a UTC container and on a laptop in another timezone. - Accuracy ledger — the admin records the day's actual outcome;
/statsscores the council against reality and/exportdumps the history as CSV. - Attendance tracking per member, counted only from their join date forward.
- Errors surface. A global handler catches slash-command and event-loop exceptions, replies to the user, and DMs the admin the traceback.
- 72 tests across the domain logic, every repository, the scheduler jobs and the trading calendar.
main.py ensure_schema() -> build_bot() -> run
config.py environment + schedule constants
src/
├── bot/
│ ├── client.py bot construction, view registration, slash-command sync
│ ├── views.py PredictionView / ConfidenceView / ResultView
│ ├── formatting.py question text, onboarding brief, consensus embed
│ └── handlers/ admin · membership · lifecycle · submission · errors
├── domain/ <- pure functions: no I/O, no Discord
│ ├── aggregator.py votes -> counts, percentages, average confidence
│ ├── classifier.py aggregate -> HIGH / MODERATE / NO_SIGNAL
│ └── stats.py accuracy and attendance
├── db/
│ ├── schema.sql members · predictions · results · daily_aggregates · skips · broadcast_log
│ ├── connection.py context manager, WAL, foreign keys on
│ ├── migrations.py idempotent schema + additive column migrations
│ └── repositories/ one module per table, dataclasses out
├── scheduler/
│ ├── runner.py AsyncIOScheduler pinned to ET
│ ├── jobs.py broadcast · consensus · session-end reminder
│ └── calendar.py NYSE holidays + is_trading_day()
├── services/
│ ├── broadcaster.py DM fan-out with delivery log and Forbidden handling
│ └── session_state.py is the submission window open right now?
└── utils/timezones.py ET helpers
The layering is the point: src/domain/ knows nothing about Discord or SQLite. It takes dataclasses and returns dataclasses, which is why the interesting logic is testable without mocking a single Discord API call. Handlers stay thin, and the scheduler jobs are the only place where I/O, persistence and domain logic meet.
Deeper notes on the data model, failure modes and design trade-offs: docs/architecture.md.
Stack: Python 3.12+ · discord.py 2.x · APScheduler · SQLite (WAL) · pytz
git clone https://github.com/SaiStyles/Wisdom.git
cd Wisdom
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env # then fill it in — see below
python scripts/seed_members.py
python main.pyStartup is healthy when the log reads:
DB schema ready.
Logged in as <YourBot> (id=...)
Synced 9 slash commands to guild ...
Scheduler started.
| Variable | Meaning |
|---|---|
BOT_TOKEN |
Discord bot token from the Developer Portal. Never commit it. |
ADMIN_DISCORD_ID |
Your Discord user ID — the only account admin commands answer to. |
GUILD_ID |
The server slash commands are synced to. |
GROUP_CHANNEL_ID |
Channel the consensus embed is posted to. |
DB_PATH |
SQLite file path. Defaults to data/council.db. |
TEST_MODE |
true unlocks three /test_* commands and bypasses the trading-day and time-window gates, so the full cycle can be exercised on a Sunday afternoon. false in production. |
Enable the Server Members Intent. Message Content Intent is not needed — this bot never reads message content. Invite with scopes bot + applications.commands and permissions Send Messages, Embed Links, Read Message History.
| Command | Who | What |
|---|---|---|
/status |
roster | Have I submitted today? |
/welcome |
roster | The full onboarding brief plus today's status. |
/add_member |
admin | Add a member via the user picker; auto-DMs the welcome text. |
/remove_member |
admin | Deactivate a member — a soft delete, so history survives. |
/add_result |
admin | Record today's actual outcome. |
/stats |
admin | Council accuracy across all signal days. |
/member_stats |
admin | Per-member attendance. |
/skip_today |
admin | Cancel the cycle for today (FOMC, half-day, a holiday the calendar missed). |
/export |
admin | Full result history as CSV. |
/test_broadcast · /test_consensus · /test_reset |
admin, TEST_MODE only |
Fire a job on demand; clear today's rows. |
Admin replies are always ephemeral.
pip install -r requirements-dev.txt
python -m pytest tests/ -q72 tests, about three seconds, no network and no Discord connection. Each runs against a fresh temporary SQLite file, so the suite is order-independent and leaves nothing behind. Scheduler jobs are exercised against a stub bot to verify delivery, deduplication and the trading-day gate.
The manual QA script for the parts only a human with a mouse can check — button flows, DM permissions, embed rendering — is in docs/testing.md.
The bot has to be running at 09:20 ET on a weekday or the day is simply lost. It ships with a Procfile and runtime.txt for Procfile-aware hosts, a run.bat auto-restart launcher for self-hosting on Windows, and a systemd unit in deploy/ for a Linux box. Options and trade-offs: docs/deployment.md.
- The signal is unvalidated. The accuracy ledger exists precisely because the premise hasn't been proven. A council that hasn't accumulated a few hundred trading days hasn't said anything statistically meaningful.
- Crowd wisdom needs a crowd. Below roughly ten diverse members the turnout gate mostly returns NO_SIGNAL — correct behaviour, and a boring one.
- SQLite suits one process on one machine. It's the right call at this scale and the wrong one the moment there are two writers; the repository layer is deliberately thin so that dialect swap stays contained.
- The holiday calendar is hard-coded through 2026 and needs a yearly top-up in
src/scheduler/calendar.py. - This is not financial advice, and neither is anything the bot posts.
MIT — see LICENSE.