Week Three — Foundation model vs Brownian motion. Kronos on five-minute BTC.
Up next
Author
Share article
The post has been shared by 0 people.
Facebook 0
X (Twitter) 0
Pinterest 0
Mail 0

📊 Full opportunity report: Week Three — Foundation model vs Brownian motion. Kronos on five-minute BTC. on ThorstenMeyerAI.com — validation score, market gap, and execution plan.

TL;DR

A recent test compared Kronos, a foundation model trained on global crypto data, against a traditional Brownian motion model for 5-minute Bitcoin predictions. The results show no statistically significant advantage for Kronos in out-of-sample testing, challenging assumptions about modern models outperforming classical methods.

Recent testing shows that Kronos, an open-source foundation model for financial time series, does not outperform the traditional Brownian motion model in predicting 5-minute Bitcoin price movements during out-of-sample testing.

Over the past two weeks, a researcher conducted a rigorous, open-source evaluation comparing Kronos-small, a foundation model trained on 45 global crypto exchanges, against a geometric Brownian motion baseline for predicting BTC’s 5-minute close direction. The test involved analyzing 497 trades, reconstructing market context, and scoring each model’s probability predictions using established metrics like Brier score and log-loss.

The results showed that Kronos’s predictive performance was statistically indistinguishable from Brownian motion on out-of-sample data, with a negligible Brier score difference of 0.0011 over 249 trades. In fact, Brownian motion slightly outperformed Kronos in the out-of-sample test, contradicting expectations that modern learned models would provide a significant edge in such short-term predictions.

According to the researcher, this outcome suggests that, at least for the specific horizon and data set tested, advanced foundation models may not yet deliver a practical advantage over classical stochastic models for high-frequency crypto trading signals.

Polybot Week 3 — Kronos vs Brownian — Thorsten Meyer AI
KRONOS
● RESEARCH SERIES / MAY 2026
THORSTEN MEYER AI · POLYBOT · WEEK 3
POLYBOT · WEEK 3
KRONOS vs BROWNIAN
Research Series · Foundation Model vs Classical Baseline · 2026-05-17

Foundation model
vs Brownian motion.
Kronos on five-minute BTC.

A modern learned model just lost to math from 1900. On 497 paired trades. Stage 2 is not happening.
Polybot’s fair-value strategy uses a 1900s geometric Brownian model to price 5-minute BTC outcomes. The natural follow-up after two weeks of negative parametric results: would a modern learned model trained on millions of real candles do better? The credible candidate: Kronos — open-source MIT-licensed foundation model, 25,000+ GitHub stars, AAAI 2026, four sizes from 4M to 499M parameters, trained on candles from 45 global exchanges. Test design: 497 paired (FILL→SETTLE) trades, Brownian baseline reconstructed line-for-line, Kronos-small (24.7M params) sampled with 16 forecast paths, scored on Brier + log-loss + hypothetical P&L, chronologically split for out-of-sample discipline. On 249 out-of-sample trades: Brownian 0.188 Brier vs Kronos 0.189 Brier. Gap 0.0011. Statistically indistinguishable. Stage 2 is not happening. But the paradox is more interesting than the verdict: when used as a directional signal Kronos fires 28% less often and wins 60.7% vs Brownian’s 49.1% — slightly better trader on hypothetical P&L, even while systematically over-confident in the tails (predicts 2.4% chance → actual 20.4% win; predicts 84% → actual 69.6%). The negative result is the answer. The methodology is what gets published.
Thorsten Meyer AI Polybot · Week 3 MIT-licensed · methodology public Research article · ~2,200 words Forezai · Polybot
This is not financial advice. Nothing in this article should be used to inform real trading decisions. The bot trades simulated money. If you build something like it and run it with real funds, the most likely outcome — by a wide margin — is that you lose those funds. That holds whether you use a Brownian model, a 100-million-parameter foundation model, or any other forecaster.
497
Paired (FILL→SETTLE) trades
all BTC · 5-min Up/Down markets
0.0011
Out-of-sample Brier-score gap
249 trades · statistically indistinguishable
2×
Kronos log-loss vs Brownian
signature of confident wrong predictions
+$538 / +$465
Hypothetical Kronos vs Brownian P&L
the paradox · 60.7% vs 49.1% win rates
POLYBOT WEEK 3· KRONOS-SMALL · 24.7M PARAMS· BROWNIAN BASELINE· 497 PAIRED TRADES · BTC· POLYMARKET 5-MIN UP/DOWN· BRIER 0.193 / 0.211 / 0.213· LOG-LOSS 0.567 / 0.604 / 1.080· OUT-OF-SAMPLE 0.188 vs 0.189· GAP 0.0011 · INDISTINGUISHABLE· STAGE 2 NOT HAPPENING· KRONOS BETTER TRADER · WORSE FORECASTER· 60.7% vs 49.1% WIN RATE· TAILS: 2.4% → 20.4% · 84% → 69.6%· POLYBOT MIT· KRONOS MIT· AAAI 2026 PAPER · 25K+ STARS· 11 MIN MAC M-SERIES · MPS BACKEND· 1,300 LINES OF PYTHON· RESEARCH_PIPELINE.MD PUBLIC· SAME GAUNTLET · DIFFERENT MODEL· POLYBOT WEEK 3· KRONOS-SMALL · 24.7M PARAMS· BROWNIAN BASELINE· 497 PAIRED TRADES · BTC· POLYMARKET 5-MIN UP/DOWN· BRIER 0.193 / 0.211 / 0.213· LOG-LOSS 0.567 / 0.604 / 1.080· OUT-OF-SAMPLE 0.188 vs 0.189· GAP 0.0011 · INDISTINGUISHABLE· STAGE 2 NOT HAPPENING· KRONOS BETTER TRADER · WORSE FORECASTER· 60.7% vs 49.1% WIN RATE· TAILS: 2.4% → 20.4% · 84% → 69.6%· POLYBOT MIT· KRONOS MIT· AAAI 2026 PAPER · 25K+ STARS· 11 MIN MAC M-SERIES · MPS BACKEND· 1,300 LINES OF PYTHON· RESEARCH_PIPELINE.MD PUBLIC· SAME GAUNTLET · DIFFERENT MODEL·
FIG. 01 — THE TEST PIPELINE
Five steps · for every paired (FILL → SETTLE) trade in the running session
~1,300 lines of Python · 11 minutes on Mac M-series with PyTorch MPS · methodology public, specific numbers local
1
Reconstruct OHLCV context of the 60 minutes leading up to fire-time. Pull from the bot’s local Binance recording where available; fall back to Binance’s public klines API otherwise. Cache to parquet so re-runs cost nothing.
2
Recompute the Brownian baseline in Python — a line-for-line port of the bot’s own fairValuePUp(spot, openPrice, secondsLeftFrac, windowVol) formula. Matches scipy.stats.norm.cdf to three decimal places.
3
Read off the market-implied probability from the FILL price — what Polymarket’s order book thought the side was worth at the moment of fire. The market’s view as a reference point.
4
Run Kronos-small (24.7M parameters) on the OHLCV context · sample 16 forecast paths to the window’s end · count the fraction in which the underlying closes above the open price. That fraction is Kronos’s predicted p(Up).
5
Record (p_brownian, p_market, p_kronos, actual_outcome, P&L). Score on Brier + log-loss + hypothetical P&L. Sort chronologically · split into first/second half · report on both halves separately.
The discipline that matters: if a model wins on the first half but ties or loses on the second, that’s the curve-fit-in-slow-motion pattern the previous two articles named, and it doesn’t count as edge. The whole pipeline is reproducible from docs/RESEARCH_PIPELINE.md. Any future candidate model gets a sibling directory in research//, reuses the same Brownian baseline, the same trade-log loader, the same OHLCV fetcher, the same metrics, the same out-of-sample split. Same gauntlet, different model, same discipline.
FIG. 02 — FULL-SAMPLE SCORING · 497 PAIRED TRADES
Three models · two probability-scoring metrics
Brier score and log-loss · the standard scoring rules for probability forecasts · lower is better
Model
Brier ↓
Log-loss ↓
BrownianGeometric Brownian motion · the 1900s baseline
0.193
0.567
Market-impliedPolymarket order book at FILL · reference
0.211
0.604
Kronos24.7M-param foundation model · 16 sampled forecast paths
0.213
1.080
Kronos’s log-loss is roughly twice Brownian’s — the signature of a model that makes confident, wrong predictions in the tails. Polymarket’s order book sits between the two, reasonably calibrated, slightly worse than the bot’s Brownian and slightly better than the foundation model. The 100-year-old math beat the 24.7M-parameter foundation model on both probability-scoring metrics.
FIG. 03 — OUT-OF-SAMPLE VERDICT · 249-TRADE TEST HALF
Chronologically-separated · never seen by tuning
The verdict the test was designed to deliver · noise band of repeated runs with different sampling seeds
Brownian · 249-trade test half
0.188
Brier score (out-of-sample)
lower is better
Kronos · 249-trade test half
0.189
Brier score (out-of-sample)
lower is better
The gap
0.0011
Statistically indistinguishable
inside the noise band
Kronos does not beat Brownian on a held-out chronologically-separated sample. So Stage 2 is not happening.
“Stage 2” was the planned next step: wiring Kronos into Polybot as a live strategy if Stage 1 produced a clear signal. The case is not earned by this data. For 5-minute BTC at the horizons the bot trades, the open Kronos-small checkpoint does not. Stop. The next candidate model — Chronos · TimesFM · Lag-Llama · a Kronos finetune on 5-min crypto · something else — goes through the same gauntlet. Most will fail it. That is the gauntlet doing its job.
FIG. 04 — THE PARADOX · BETTER TRADER vs WORSE FORECASTER
By operational standards Kronos wins · by probabilistic standards Kronos loses
The hypothetical-P&L counterfactual replays the same data through “what if Polybot fired on each model’s probability”
Operational view · Kronos as the better trader
Kronos fires less · wins more · nets slightly more.
Hypothetical fires
201
Brownian fires (reference)
279
Win rate (Kronos)
60.7%
Win rate (Brownian)
49.1%
Hypothetical net P&L (Kronos)
+$538
Hypothetical net P&L (Brownian)
+$465
Fires ~28% less often and wins more reliably when it does. If you use Kronos as a directional signal in a broader system that does its own sizing — closer to how TradingAgents uses analyst outputs — the directional accuracy might still be useful.
Probabilistic view · Kronos as the worse forecaster
Systematically over-confident in the tails.
Kronos predicts
2.4%
Trades actually win
20.4%
Kronos predicts
84%
Trades actually win
69.6%
Log-loss vs Brownian
~2× worse
Brier (full sample)
0.213 vs 0.193
If you are building a fully-probabilistic system where the probability feeds an expected-value calculation against the market’s implied price — which is what Polybot does — calibration is everything, and Kronos’s calibration is bad enough to disqualify it. It thinks it knows more than it does at both ends.
Both interpretations are honest. Neither earns the model a place in Polybot. One of them might earn it a place, later, in TradingAgents — as a 5th analyst voice that votes on direction without being trusted for calibrated odds. That experiment is not what this week tested; it is a separate hypothesis for a separate week.
FIG. 05 — WEEK FOUR · THREE POSSIBLE THREADS
Each is a separate article · the pattern across them is the same
Honest measurement · out-of-sample discipline · no rescue narratives when something doesn’t work
1
A second-tier candidate model · Amazon’s Chronos
Same general shape as Kronos · different training corpus · also open-source. Running it through the exact same gauntlet would say whether the negative result is specific to Kronos or generalises to learned models in this regime.
Generalisation test
2
Kronos with a finetune on 5-min crypto data
The Kronos repo ships a finetuning pipeline. Taking the open Kronos-base checkpoint, finetuning on the bot’s own recorded BTC tick history, re-testing. Isolates “is the pretrained distribution wrong for crypto?” from “is the architecture wrong for this horizon?”
Architecture vs distribution
3
A live-trading update on Polybot
The fleet has been running paper trades continuously across these three weeks. A fresh aggregate-P&L view, with the same calibration-style analysis applied to live performance rather than historical replay, is overdue.
Status reset
The contract is “same gauntlet, different model, same discipline.” Specific numbers stay local. Methodology is public on the repo’s docs/RESEARCH_PIPELINE.md. Publishing reproducible parameter recipes for strategies that might be marginally profitable encourages people to copy them with real money, and the prior on real-money outcomes when copying retail strategies is “they lose.” Publishing the methodology lets the next person test their own model honestly without inheriting any of mine.
By probabilistic standards · Kronos is a worse forecaster. By operational standards · Kronos is the better trader. Both interpretations are honest. Neither earns the model a place in Polybot. One of them might earn it a place, later, in TradingAgents.
Thorsten Meyer AI · Week 3 · Foundation Model vs Brownian Motion
Source dossier & methodology notes
  • Kronos — Open-source MIT-licensed foundation model for financial time series · 25,000+ stars on GitHub · AAAI 2026 paper · four model sizes (4M · 24.7M open small · 102M open · 499M closed) · trained on candles from 45 global exchanges · authors are explicit it is a research model, not a trading system · github.com/shiyu-coder/Kronos
  • Polybot — Open-source paper-trading bot · MIT-licensed · runs against Polymarket 5-minute Up/Down crypto markets · fair-value strategy uses geometric Brownian motion · two prior weeks of published research established that most parametric edges are mechanical artefacts
  • Polymarket 5-min Up/Down — Prediction-market windows on BTC, ETH, and other crypto · the test session: 497 paired (FILL → SETTLE) trades, all BTC · order-book at FILL serves as market-implied probability reference
  • Test design — Run for every paired trade · reconstruct 60-minute OHLCV context · recompute Brownian baseline (line-for-line port of fairValuePUp(spot, openPrice, secondsLeftFrac, windowVol)) · read market-implied probability · sample 16 Kronos forecast paths · record (p_brownian, p_market, p_kronos, actual_outcome, P&L)
  • Scoring rules — Brier score (mean squared error of probability vs actual outcome) · log-loss (penalises overconfidence) · hypothetical P&L (counterfactual if Polybot had fired on each model’s probability with the same edge-margin and risk gates) · all reported on first/second half separately
  • Full-sample (497 trades) — Brier: Brownian 0.193 · Market-implied 0.211 · Kronos 0.213 · Log-loss: Brownian 0.567 · Market-implied 0.604 · Kronos 1.080 · Kronos log-loss ~2× Brownian — confident-wrong-in-tails signature
  • Out-of-sample (249 trades, test half) — Brownian 0.188 · Kronos 0.189 · gap 0.0011 · well inside the noise band of repeated runs with different Kronos sampling seeds · statistically indistinguishable
  • Counterfactual P&L — Brownian: 279 fires / 49.1% win rate / +$465 net · Kronos: 201 fires / 60.7% win rate / +$538 net · Kronos fires ~28% less often, wins more reliably, nets slightly more by operational standards
  • Calibration failures — Kronos predicts 2.4% chance → trades win 20.4% · Kronos predicts 84% chance → trades win 69.6% · systematic overconfidence at both tails
  • Compute — Mac M-series with PyTorch MPS backend · ~1,300 lines of Python · 11 minutes clock-time · cached to parquet so re-runs cost nothing
  • Public methodology — docs/RESEARCH_PIPELINE.md on the project repo · same gauntlet runs any future candidate forecast model (Chronos · TimesFM · Lag-Llama · Kronos finetune · other) · sibling directories research//
  • What this does NOT prove — Not that Kronos is bad (one checkpoint · one horizon · one market) · not that Brownian is good (just not worse at this task; week-2 collapsed at higher sample) · not anything about Stages 2 or 3 of the broader pipeline · negative-Stage-1 kills this candidate at this horizon · the gauntlet does its job
  • Week 4 candidate threads — (a) Amazon’s Chronos generalisation test · (b) Kronos finetune isolating architecture vs distribution · (c) Polybot live-trading aggregate-P&L update with calibration-style analysis
Colophon

Set in Source Serif 4 (display, italic accent), EB Garamond (body), IBM Plex Sans (UI labels), IBM Plex Mono (mastheads, ticker, tags). Paper-cool gray-cream #e6e7e4.

Chromatic register: structural-slate dominant (methodology + out-of-sample-discipline analysis), empirical-clay for the data forensics, labor-rose for the probabilistic-failure framing and disclaimers, transition-bronze for the operational paradox, alternative-sage for the open-methodology positive signal and the better-trader half of the paradox.

Key frame:

POLYBOT WEEK 3 KRONOS-SMALL · 24.7M BROWNIAN BASELINE OUT-OF-SAMPLE 0.0011 THE PARADOX BETTER TRADER WORSE FORECASTER STAGE 2 NOT HAPPENING SAME GAUNTLET DISCIPLINE HONEST MEASUREMENTS

Implications for AI-based Crypto Trading Strategies

This testing challenges the assumption that modern, data-driven foundation models automatically outperform traditional mathematical models like Brownian motion in short-term crypto prediction. While Kronos incorporates extensive global exchange data, its inability to beat the classical model in this context indicates that current AI models may require further development before offering consistent, actionable edge in high-frequency trading.

For traders and algorithm developers, this suggests caution in deploying foundation models as standalone predictive tools for short-term crypto trades. It also emphasizes the importance of rigorous out-of-sample testing to validate any claimed advantage of advanced models over simpler, well-understood stochastic processes.

CafePress Bitcoin 5 1" Round Mini Button Pin

CafePress Bitcoin 5 1" Round Mini Button Pin

MEASUREMENTS: 1" round.

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Previous Attempts and Expectations for Foundation Models

Historically, traders and researchers have relied on geometric Brownian motion as a foundational assumption in modeling asset prices, including cryptocurrencies. Recent advances in foundation models like Kronos, trained on massive datasets from multiple exchanges, raised hopes that such models could surpass traditional methods in short-term prediction accuracy.

Prior experiments, including the author’s two-week paper-trading bot testing, indicated that most ‘edges’ found by automated strategies were mechanical artifacts that did not survive out-of-sample validation. The question then became whether a modern, learned model could provide a genuine predictive advantage in the volatile, high-frequency crypto environment.

Previous literature suggests that while foundation models excel in natural language and image domains, their efficacy in financial time series remains uncertain, especially at very short horizons like five minutes.

“The results show no statistically significant outperformance of Kronos over the Brownian baseline in out-of-sample tests, challenging expectations for modern models.”

— Thorsten Meyer, researcher

Trading Chart (Set of 5) Posters, 350 GSM Candlestick Pattern Cheat Sheet, Trade Setup Kit for Stock, Forex and Crypto Market (30 x 21 CM, Unframed)

Trading Chart (Set of 5) Posters, 350 GSM Candlestick Pattern Cheat Sheet, Trade Setup Kit for Stock, Forex and Crypto Market (30 x 21 CM, Unframed)

Complete Trading Chart Guide: Master market analysis with this detailed Candlestick Pattern Cheat Sheet featuring essential bullish, bearish,…

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Unanswered Questions About Model Performance

It remains unclear whether different model configurations, larger training datasets, or alternative prediction horizons might yield better results. Additionally, the performance in live trading conditions, with transaction costs and slippage, has not been tested.

Further research is needed to determine if foundation models can be optimized for short-term crypto trading or if their advantages are limited to longer-term or different market contexts.

Developing High-Frequency Trading Systems: Learn how to implement high-frequency trading from scratch with C++ or Java basics

Developing High-Frequency Trading Systems: Learn how to implement high-frequency trading from scratch with C++ or Java basics

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Next Steps for Evaluating Foundation Models in Crypto

The researcher plans to extend testing to other market conditions, different horizons, and larger models. There is also interest in integrating foundation models into live trading simulations with transaction cost considerations.

Meanwhile, the current results suggest that traders should remain cautious about over-reliance on AI models without rigorous out-of-sample validation, especially for high-frequency strategies.

Bitcoin 1$ Million Price Prediction Cryptocurrency BTC T-Shirt

Bitcoin 1$ Million Price Prediction Cryptocurrency BTC T-Shirt

Lightweight, Classic fit, Double-needle sleeve and bottom hem

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Key Questions

Does this mean foundation models are useless for crypto trading?

Not necessarily. The current test shows they did not outperform a classical model in this specific short-term prediction task. Further research and different configurations may yield different results.

Could larger or differently trained models perform better?

Potentially. The tested model was one size among several, and different training data or architectures might improve performance. This remains an open area for investigation.

Is this testing relevant to live trading?

The test was conducted offline and does not account for transaction costs, slippage, or market impact. Live trading conditions could influence the effectiveness of any model.

Will future research change these results?

Future studies with larger datasets, different horizons, or more sophisticated models could produce different outcomes. The current findings are specific to this setup.

Source: ThorstenMeyerAI.com

You May Also Like
Opus 4.8 Lands, and the Quiet Headline Is Honesty

Opus 4.8 Lands, and the Quiet Headline Is Honesty

Anthropic releases Claude Opus 4.8, highlighting improved honesty and safety measures alongside performance gains in benchmarks, amid a challenging month.
The referral. How AI search severs the content-for-traffic contract that funded the open web.

The referral. How AI search severs the content-for-traffic contract that funded the open web.

AI search now answers queries directly, cutting off publishers’ referral traffic and revenue, fundamentally shifting the digital publishing economy.
Jack Clark Says It Out Loud — Reading the Co-Founder’s 60%/2028 Estimate on Automated AI R&D

Jack Clark Says It Out Loud — Reading the Co-Founder’s 60%/2028 Estimate on Automated AI R&D

Anthropic’s Jack Clark states there is a 60%+ probability that AI systems capable of autonomously building successors will emerge by 2028, signaling a major policy forecast.
One Video In, a Whole Publishing Kit Out — Without the Cloud

One Video In, a Whole Publishing Kit Out — Without the Cloud

New local-first workflow transforms a single video into a full suite of publishing assets offline, saving time, costs, and enhancing privacy.