ToneBot

The system follows the same five-step loop every trading day. It does not improvise and it does not act on tips.

1. Read the market

   Each area of the market is scored on how strongly, and how steadily, it has outperformed the overall market. The system also reads conditions: whether the market is trending or choppy, whether interest rates are a tailwind or a headwind, and whether fear is rising.

2. Select only from leaders

   New positions come only from leading areas, and only from investments already in their own uptrend with enough trading volume to enter and exit cleanly.

3. Size by risk

   Position size depends on how much a position could lose, not on how attractive it appears. More volatile holdings receive smaller allocations; leveraged funds are treated as the larger risk they carry.

4. Exit by rule

   Every position has an exit plan before it is bought: a protective stop that rises as the position gains, and a time limit to prove itself. Gains are allowed to run; losses are cut early.

5. Review

   After the close the system grades its own decisions, compares actual fills with the plan, and records what it would change — without changing anything until a scheduled checkpoint.

A system versus a human trader

Neither is better at everything. The system is built to do what people do badly, and its known weaknesses are watched rather than hidden.

Backtests can also flatter a system. Ours have been corrected twice so far: removing investments added to the list only after they had risen, and adding the real costs of leveraged funds to historical rebuilds.

ToneBot studies every week, but its trading rules change only at scheduled checkpoints. That separation keeps the public test honest while the research continues.

Findings

• 2026-10-10
  Risk

  Honest drawdowns are bigger than they first look

  Our older backtests only counted a loss when a trade closed. Counting every open position at every day's close shows the worst drop of the current rules in some past crashes was larger than our own 35% limit. We now measure everything that way, and no change can be promoted unless it stays inside the limit across every historical period we test.

• 2026-10-10
  Risk

  The fix for big drops is sizing, not prediction

  Across published research and our own tests, the changes that cut the worst drops were about HOW MUCH to hold in certain conditions, not about predicting the next move. One combination of sizing rules kept every tested period inside our limit and did better in most of them. It still has to pass our stress test before it is a candidate for January.

• 2026-10-10
  Forecasting

  Nobody (including us) can call one stock's next month

  We tested 'look-alike' forecasts on 7,732 past cases across 12 well-known stocks and funds. Guessing whether a single stock would be up or down 20 days later was no better than a coin flip. The forecast RANGES, though, were well calibrated: about 79% of outcomes landed inside our 80% range. So the bot uses forecasts to size risk, never to pick winners.

• 2026-10-10
  Strategy

  Bolder is not better

  We tested five aggressive or contrarian alternatives side by side (leveraged trend-following, all-in on the hottest sector, earnings-gap buying, buying fear, fading dips). None beat the current approach on recent years, and the most aggressive one nearly wiped out in the 2000-02 crash. The edge is in discipline, not in swinging harder.

• 2026-10-10
  Reliability

  A trading bot must survive its own failures

  We audited every way the system could go quiet or need a human: data sources, the broker login, a sleeping computer, alert channels. Each hole got a full fix — stale prices block new buys, alerts fall back to a phone push, two machines can never trade the same account at once, and a backup machine can take over.

• 2026-10-10
  Learning

  It learns every week, but changes only at checkpoints

  During the test the trading rules are frozen so the result is honest. The bot still records what it WOULD change, grades its own forecasts and trades, runs experiments on years it has never seen and reads new research every Saturday. Only scheduled checkpoints can switch anything on.

Where attention went

Every time the system or its research desk reads, measures or decides something, it is logged. Activity over the last 7 days, by area:

Share of the work: the automated system 273, the research desk 60, the founder 131.

The pass marks were written before the test began, so they cannot be moved. On January 18, 2027 each is scored publicly. Real capital is committed only if all nine pass, and then at a quarter of normal size, scaling up over several months.

The weekly scoreboard begins once the test is under way.

Practice comparison since inception: strongest strategy +3.3%, weakest -2.2%.

Safety settings are always active and are not altered during the test.

Topics studied by the research desk, newest first. Published research is a source of ideas; nothing is adopted without passing our own tests.

Recent papers under review

• Continuous Timing Signals for Growth-Defensive Style Allocation: Factor Attribution, Risk Matching, and Out-of-Sample Evidence
• Forecast-to-Fill: Benchmark-Neutral Alpha and Billion-Dollar Capacity in Gold Futures (2015-2025)
• End-to-End Parametric Portfolio Policies for Cross-Asset Futures Timing: When Do AI Models Beat Simple Rules?
• Observable Matrix Dynamics of Stocks
• (Non-Parametric) Bootstrap Robust Optimization for Portfolios and Trading Strategies
• Deep Hedging with Reinforcement Learning: A Practical Framework for Option Risk Management
• Regime-Conditional Distributional Comparison of Trading Strategies: A GAMLSS/ZAGA Framework Applied to the S&P 500
• Refining and Robust Backtesting of A Century of Profitable Industry Trends

What this system would take to build by hand, set against what it has actually cost.

The system has cost roughly 0.2% of what a hand-built equivalent would. It was designed and directed by its founder — every rule, idea and approval above — and written with an AI assistant (Claude), which is where nearly all of the difference comes from. Founder time is a floor: most time spent in planning conversations leaves no trace in the logs.

Improvements to the system, newest first. Strategy changes occur only at checkpoints; most entries concern measurement, safety and reliability.

Disclosures

This paper describes a system under test. It is not an offer, a solicitation or investment advice, and it should not be used to make investment decisions. Historical and simulated results have inherent limitations and do not guarantee future results. Live results may differ materially. The system's specific rules, parameters, holdings and investment universe are proprietary and are intentionally withheld.