Aviator Predictor

An offline Python/Streamlit predictor and data logger for Aviator-style crash-game rounds. It exists to make one call - stake or skip - from your logged rounds plus the game's verified 97% RTP maths, and it tells you honestly when it doesn't have enough data yet.

No betting automation. No casino connection. No guaranteed predictions. Every round of a provably fair crash game is independent; this tool gives probability estimates, never certainties.

-----> https://aviator-odds.streamlit.app/

Features

Predict tab (the point of the app)

The main page is intentionally sparse - a slim header (title + a status chip like 60 rounds · THEORY mode), one honest disclaimer line, then the hero:

• PREDICT - one big circular button. Press it and the result circle shows the suggested odd, whether the minute is a Golden window or Outside window, plus P(reach) … · pressed HH:MM:SS; a quiet Reset pill clears it for the next round. Below the hero: the round selector (Auto from the clock, or any minute 00-59) and a 4-pill strip showing the next minutes with their odds.
• Always an odd - every minute gets a suggestion in 1.00-2.00x: minutes in your golden windows and data-validated minutes get a data-based odd (usually the cautious 1.00-1.45x band), and minutes outside the golden windows get an advanced odd taken from the enrichment CSV, clearly flagged OUTSIDE GOLDEN WINDOW instead of being skipped. The same odd never shows twice in a row - if a repeat would appear (presses, strip pills or plan rows) it is nudged one 0.05 grid step inside its band and the Details card says so.
• Everything else sits behind collapsed expanders so the page stays uncluttered:
  • Details - the reasoning card (alert, confidence, basis, odd source, metric cards, comparison tables), the data-quality metrics (timestamped / in-window / needed-for-data-backed / mode) and the progress bar.
  • Minute plan & windows - the 60-minute plan, window health tables with Fisher exact + Mann-Whitney U verdicts (p hidden below n = 30), editable golden windows (defaults 03-04, 20-22, 29-31, 40-42, 45-47, 50-52, 57-59), data-favoured minutes gated by Wilson intervals, and the cash-out target table (every target -3% EV, 97% RTP).
  • Sources & safety - Spribe's RTP 97%, the published provably fair formula P(reach x) = 0.97/x with the FairPlay Audit verification, and the responsible-gambling contacts (Botswana 1144 etc.).
• Sidebar - a collapsed Status expander (rounds stored, data file path with copy button, last refresh time), the developer credit: Developed by Karabo Kosi with icon links (phone, Gmail, website, LinkedIn, YouTube, Instagram, TikTok) and the phone + email in text, and a collapsed Rules & stake strategy expander right under the contacts - the play rhythm (press PREDICT, sit out the next round, then enter every second round: 1, 3, 5 …), the risk ↔ target rule (risk low → target high, risk high → target low) and the doubling stake ladder (P10 → P20 → double after each loss until a win, then back to P10), with the honest note that the ladder never changes the -3% edge.
• Mobile friendly - below 760px every column row stacks vertically (hero, metrics, editors), the PREDICT and odd circles shrink to 200px (150px in short landscape viewports), the minute strip wraps and the sidebar collapses into the hamburger menu.

Data tab (feeding the predictor)

Quality summary at the top, then three tidy sections:

• Add rounds (open by default) - fast live entry (type the multiplier, Enter; timestamp fills in automatically in Africa/Gaborone, UTC+02:00) and the bulk paste form (start time + seconds-per-round interval, or stamp existing untimed rows).
• Import / Export - CSV merge upload with duplicate skipping, export and blank template downloads, plus the Enrichment CSV block (upload or reload the bundled csv_data_odds/aviator_rounds.csv of odds with estimated timestamps; it drives the outside-golden-window odds. Bundled file: 334 rounds read by eye from screenshots, times estimated at ~22 s/round (±1 min, gaps between segments)).
• Danger zone (red-outlined buttons, each behind a confirm step) - Restore seed data and Clear all data.
• In-app editor - st.data_editor over the logged rounds (Apply validates every row: multiplier ≥ 1.00, valid source).

Requirements

• Python 3.11+ (developed and tested on Python 3.12, Windows)

Windows setup

cd path\to\aviator_odds
python -m venv venv
venv\Scripts\activate
pip install -r requirements.txt
streamlit run app.py

Then open http://localhost:8501 in your browser.

If you prefer not to activate the venv:

venv\Scripts\python.exe -m pip install -r requirements.txt
venv\Scripts\python.exe -m streamlit run app.py

Data

Rounds are stored in data/rounds.csv with columns timestamp,multiplier,source (sources: manual, import, seed). Rows without a timestamp (the bundled 60-round seed sample) are excluded from time-based analysis until stamped.

• Set the AVIATOR_CSV environment variable to use a different file path.
• Set AVIATOR_NO_AUTOREFRESH=1 to disable the live sidebar/timer auto-refresh.
• On first run the seed file is created automatically; Restore seed data and Clear all data are on the Data tab.

CSV format

timestamp,multiplier,source
2025-03-01T14:05:00+02:00,1.50,manual
,83.73,seed

Duplicates (same timestamp + multiplier) are skipped on merge; untimed rows are merged per-multiplier with a max-count rule.

Tests

venv\Scripts\python.exe -m pytest

Covers the statistics functions against reference values (Wilson intervals, percentiles, Fisher exact, Mann-Whitney, chi-square, BH/Bonferroni correction), the prediction engine, CSV parsing/merging/seed data, and the app itself via streamlit.testing.v1.AppTest.

Project layout

app.py          Streamlit UI (2 tabs: Predict, Data)
stats.py        Statistical helpers + prediction engine (pure functions)
datastore.py    CSV parsing, merging, seed data, time helpers
data/rounds.csv Round storage (created on first run)
tests/          pytest suite

Honest limits

• Confidence intervals widen and verdicts become unreliable on small samples - the app says Not enough data rather than guessing.
• Scanning 60 minutes at once manufactures false positives; corrected p-values (BH, Bonferroni) account for that.
• No staking system beats a negative expected value. Martingale loses faster when it loses; flat staking bleeds at the same theoretical rate.