# Kronos / Polymarket Scalping Bot — Feasibility & Risk Analysis

**Date:** May 5, 2026 | Prepared for Client

---

## Executive Summary

The YouTube video by David (TradeLikeBoss) shows a bot that uses Kronos AI to predict 5-minute crypto candle direction and trade on Polymarket's binary markets. He claims $21,000 profit from $100 in 90 days.

**My verdict: The claimed profits are almost certainly the result of overfitting and data leakage. The underlying AI model has been independently tested at only 12% direction accuracy — worse than a coin flip. However, the architecture and approach are interesting, and with modifications, we could build and test something similar.**

---

## What The Video Claims

- Built using **Kronos** (open-source time-series AI model) + **Claude Code** + custom workbook
- Trades on **Polymarket** 5-minute and 15-minute BTC/ETH up/down markets
- Dashboard called "Crow Terminal" shows predictions, confidence, equity curve
- Uses **ATR bands + ADX** as gatekeepers to filter bad trades
- Uses **Kelly Criterion** for position sizing
- Claims: 200,000 backtest trials, 15,000 bets
- Kelly-only: $12,300 profit in 3 months
- Martingale + Kelly: $23,300 profit in 90 days
- Starting capital: $100

---

## What Kronos Actually Is

**Kronos** is an open-source time-series foundation model on GitHub (22.9k stars):
- Paper accepted at AAAI 2026
- Built by Yu Shi, Zongliang Fu et al.
- Uses a Binary Spherical Quantizer (BSQ) tokenizer to convert OHLCV candlestick data into discrete tokens
- Autoregressive transformer generates future candle predictions
- Available models:
  - Kronos-mini (4.1M params) — runs on Mac
  - Kronos-small (24.7M params)
  - Kronos-base (102.3M params)
  - Kronos-large (499.2M params) — NOT released

**Can run locally** on a Mac Mini or MacBook Pro — no cloud needed for the smaller models.

---

## Red Flags — Why The Claims Don't Hold Up

### Red Flag #1: Sub-Random Direction Accuracy

**GitHub Issue #282** (May 1, 2026): Independent cross-validation test of Kronos-small across 16 top global assets (30 years daily data) showed:
- **Direction accuracy: 12% (2/16 assets correctly predicted)**
- This is SIGNIFICANTLY WORSE than random (50%)
- Strong bearish bias: model systematically predicts decline
- BTC predicted -26.6%, actual +49.8%
- ETH predicted -6.3%, actual +55.3%
- Conclusion from tester: **"Not suitable for actual trading decisions"**

### Red Flag #2: Confirmed Data Leakage

**GitHub Issue #227** (Mar 2026, confirmed via PR #234): The fine-tuning code has a **data leakage bug** — per-window normalization computes mean/std from the FULL window (including the future prediction horizon), contaminating all fine-tuned results. This means:
- Any "backtested" results using the fine-tuned model are unreliable
- The claimed paper results may be valid (pre-training uses different normalization) but fine-tuned applications are compromised
- This directly affects the video's backtest results

### Red Flag #3: Martingale + Kelly

- **Kelly Criterion** already optimizes growth rate by sizing positions based on edge
- **Martingale** (doubling after a loss) is a well-known ruin strategy
- **Combining them** is exceptionally dangerous — when the model has no real edge, Martingale accelerates losses exponentially
- Claiming $23,300 from $100 in 90 days using Martingale+Kelly is a classic scam narrative
- With no edge, you'd blow up the entire account almost certainly

### Red Flag #4: Survivorship Bias

- The video only shows the winning configuration
- Countless parameter combinations (ATR thresholds, ADX levels, Kelly fractions, lookback windows) lose money
- The "best" result is selected after seeing all outcomes — this is the definition of overfitting

### Red Flag #5: 5-Minute Markets + Fees

- Polymarket charges 0-2% per trade depending on maker/taker status
- 5-minute binary markets have tight spreads
- With marginally-better-than-random predictions, fees eat all profits
- The shorter the timeframe, the more noise dominates signal

---

## Could We Build Something Similar That Actually Works?

**Yes, but with major modifications.** Here's what a legitimate version would look like:

### What The Video Gets Right
1. **Time-series models** (like Kronos) ARE the right architecture for price prediction — not LLMs like ChatGPT/Claude
2. **Polymarket** is a legitimate prediction market with a real API
3. **ATR + ADX filtering** is sound — only trading in volatile, trending markets
4. **Backtesting** before live trading is essential
5. **Kelly Criterion** position sizing is mathematically optimal IF you have a genuine edge

### What Needs To Change
1. **Don't rely on Kronos alone** — combine multiple signals:
   - Kronos direction prediction (with known 12% accuracy)
   - RSI, MACD, Bollinger Bands
   - Order book imbalance / funding rates
   - Macro signals (DXY, VIX, BTC dominance)
   
2. **Only trade when multiple signals agree** — this is the "confluence" approach and it dramatically improves accuracy

3. **Longer timeframes** — 5-minute is too noisy. 1-hour or 4-hour candles have more signal and less noise

4. **Forward-test only** — no backtest parameter tuning. Set parameters, then run live with tiny amounts for 500+ trades

5. **Flat position sizing** — not Kelly, not Martingale. 1-2% of bankroll per trade until you've proven a real edge exists

6. **Track against random baseline** — run a coin-flip strategy in parallel. If you can't beat random, you have no edge

---

## Architecture For A Test Build

If we wanted to build and test this, here's what's needed:

### Phase 1: Data Pipeline (1-2 days)
- Polymarket API connection (Python SDK exists)
- BTC/ETH price data (5-min, 15-min, 1-hour candles)
- Kronos model running locally (Mac Mini compatible)
- ATR, ADX, RSI, MACD indicator calculations

### Phase 2: Signal Generation (1-2 days)
- Kronos prediction engine
- Multi-indicator confluence filter
- Confidence scoring (only trade when 3+ signals agree)
- Paper trading (no real money) for 2-4 weeks

### Phase 3: Backtest Engine (1-2 days)
- Historical Polymarket data
- Include fees, slippage, rejected orders
- Compare vs random baseline
- Walk-forward validation (train on period A, test on period B)

### Phase 4: Live Testing (4-8 weeks minimum)
- Start with minimum bet sizes
- Track every trade meticulously
- Compare against random baseline
- Only increase position size after proving edge exists

### Phase 5: Dashboard (1-2 days)
- Build with Python + Streamlit or similar
- Show predictions, confidence, trade history, equity curve
- Alert system for high-conviction setups

**Total build time: ~1 week**
**Total testing time: 4-8 weeks before risking real money**

---

## Costs & Requirements

| Item | Cost |
|------|------|
| Polymarket account | Free (need crypto wallet) |
| Starting capital for testing | $50-100 (minimum bets) |
| Kronos model | Free (open source, runs on Mac) |
| Claude Code | $20/month (Pro) or $100/month (Max) |
| Python + deps | Free |
| Hosting (if not local) | $0-20/month |

**Minimum to get started: $0 (paper trading) or $50-100 (live micro-testing)**

---

## Risk Assessment

| Risk | Likelihood | Impact |
|------|-----------|--------|
| Kronos accuracy is sub-random | VERY HIGH | Total loss |
| Data leakage in backtests | CONFIRMED | False confidence |
| Martingale blow-up | HIGH | Account wipeout |
| Polymarket fee drag | HIGH | Slow bleed |
| Overfitting to historical data | VERY HIGH | False profits in backtest, losses live |
| Model regime change | MEDIUM | Edge disappears |
| Polymarket withdrawal/liquidity | LOW-MEDIUM | Can't exit |

**Overall risk: VERY HIGH**

---

## The Honest Truth

The video author (David) is charismatic and the dashboard looks impressive. But:

1. **Kronos has been independently tested at 12% direction accuracy** — worse than flipping a coin
2. **The data leakage bug means backtested results are contaminated** — the $21,000 profit claim cannot be trusted
3. **Martingale + Kelly from $100 to $23,300 in 90 days** is the kind of return that would make Wall Street's best quants look like amateurs — if it were real, hedge funds would be all over it
4. **The workbook is free because the real business is selling access to the idea, not the results**

The most likely scenario: the backtest was overfitted to historical data, and the Martingale component inflated returns during a favorable period. In live forward-testing, this strategy would almost certainly lose money.

**We CAN build and test something similar**, but we should:
- Use multiple signal sources, not just Kronos
- Forward-test with paper trading for at least 500 trades before risking real money
- Compare against random baseline
- Use flat position sizing (no Martingale)
- Track all fees and slippage

If after 500 trades with real money we can't beat random by at least 5% after fees, we shut it down.

---

## Recommendation

**Interest level: Moderate.** The concept is interesting and the technology is accessible. But the specific implementation in the video is statistically unsound.

**If Gary wants to proceed:**
1. I can build the dashboard and signal engine (1 week)
2. Run paper trading for 4 weeks (no real money)
3. Compare against random baseline
4. If results show genuine edge (>5% after fees), start with minimum $50 position sizes
5. Never use Martingale. Use flat 1% position sizing until edge is proven over 1000+ trades

**If Gary DOES NOT want to proceed:**
That's also a very reasonable decision. The risk of total loss is high, and the video's claims don't hold up to scrutiny.

---

*Disclaimer: This is research and analysis, not financial advice. Polymarket trading involves significant risk of loss. Past backtested results (especially those with data leakage) are not indicative of future performance.*