The Lab

The question

Our paper bots read two plain-words labels for each market every hour. We call them market words:

• Trend words, from the ADX indicator (J. Welles Wilder's, 1978; 14 periods, the standard bands). ADX scores how strong a trend is, from 0 to 100, over the last 14 bars. It measures strength, not direction.
  • ranging when ADX is under 20;
  • leaning up/down from 20 to 30;
  • trending up/down at 30 and over.
  • The direction comes from whichever side of the indicator is stronger.
• Calm or wild, from a hidden Markov model of the coin's hourly moves (Baum and Petrie, 1966).
  • Calm hours make small moves and wild hours make big ones.
  • The model learns where the line falls from the coin's own past.
  • Each hour it gives a chance of each, not a hard label.
  • The models page works an example through.

The idea is a gate: a bot only opens new trades under the words it is allowed to trade in. Exits always run. A gate is worth building only if the words actually warn of losses. For example, short trades opened when the market reads "trending up" should do worse.

So the question is: do trades opened under a word do worse than trades opened under the others?

How we looked

• The data: 195 readings from 7 October 2026 05:00 to 14 October 2026 08:33 UTC. Most are hourly; a few extra fall at fixed times of day. Alongside them are every fill and the minute-by-minute positions of the paper bots: linear and spot bots on an exchange's demo and on a testnet. One bot on an inverse market was left out, because the readings don't price it. Nothing was guessed.
• Entries: every trade a bot opened was judged by where the price was 4 hours and 24 hours later. The move is counted from the trade's side, as a percent of the price. A short opened at 100 with the price at 98 a day later counts as +2%; the same move counts as −2% for a long. That is what a gate would have kept out.
• Hours: each bot's result in each hour, as a percent of its money, grouped by the word in force at the time. This is context: what holding through that word cost.

What it showed

Trades opened under each trend word, average move after 24 hours, with the share that lost:

Under calm or wild:

In brackets: the number of trades with a 24-hour result.

• The up words came this week, and they marked tops, not rallies. Shorts opened under "trending up" all made money a day later, and longs opened under it nearly all lost. A gate that blocked shorts in an up trend would have blocked the winning trades.
• The down words marked turns too. Longs opened under "trending down" did better than longs under any other trend word.
• After this week's wild hours the price bounced. Longs opened in them did well and shorts did badly. That describes this week's wild hours, not wild hours in general.
• The hourly results swing hardest where the hours are fewest. Some paper bots run very high leverage on purpose, as a stress test, and that magnifies every hour.

Update, 27 October 2026: where the idea comes from

Added after this page was published. The text above is left as it was.

• Turning a strategy on and off by the market's state is an old idea. J. Welles Wilder built the ADX in 1978 to tell a trending market from a ranging one, so that a trend-following rule could be used only in a trend.
• The ADX, step by step.
  1. Each bar, +DI measures how far the price pushed above the last bar's high, and −DI how far it pushed below the last bar's low. Both are a share of the bar's usual range, averaged over 14 bars.
  2. DX is the gap between them divided by their total, times 100. With +DI at 30 and −DI at 10: 20 ÷ 40 × 100 = 50.
  3. The ADX is DX averaged again over 14 bars.
• Hidden Markov models come from Leonard Baum and Ted Petrie's work in 1966. James Hamilton brought switching between hidden states into economics in 1989, to tell recessions from expansions. "Regime switching" in markets follows his model.
• The hidden Markov model, step by step. The model says each hour is calm or wild, but nobody sees which: we see only the move. Say calm hours move about 0.5% and wild hours about 2%. The last reading gave 90% calm. Then an hour moves 2%.
  1. How likely is a 2% move under each? Under calm it is 4 typical moves out, which the bell curve makes very rare. Under wild it is 1 typical move out, which is ordinary. Worked out, a wild hour is about 450 times as likely to give it.
  2. Weigh that by the chances before: calm 0.9 × 1 = 0.9, against wild 0.1 × 450 = 45.
  3. The chance it is now wild: 45 ÷ (45 + 0.9) = about 98%.
  4. Before the next hour, the model also allows for the chance of a switch, which it learns from the coin's past.
• The catch. Both read the past.
  • The ADX is an average of an average over 14 bars. So it is highest after a move has already run. That is where this page found the strong words: near turns.
  • The model's chance of "wild" rises after big moves have happened. It describes the hours just gone, not the next ones.
• Theirs and ours. The ADX and the hidden Markov model are theirs, credited below. The cut-offs at 20 and 30 are conventions traders use. The market words, the gate and this way of judging it are ours.

Update, 30 October 2026: the fifth look

Added after this page was published. The text above is left as it was.

The same question, asked again of the paper bots' trades from 22 October 2026, 15:36 UTC, to 30 October: 632 readings, none left out. Every trade in this window was a long; no shorts were opened, so there is no short side to show. Two looks between this one and the second, kept in our notes, came to the same answer.

Longs opened under each word, average move after, signed so a rise is a gain for a long, with the share that lost:

In brackets: the trades with a 24-hour result so far.

The bots' hours under each word. A bot-hour is one bot running for one hour; the average is its result in that hour as a share of the bot's money. 1,443 calm bot-hours averaging −0.13% means a bot lost about an eighth of a percent of its money in a typical calm hour.

• "Trending down", the word a gate would hold longs back under, did not mark worse trades. Its longs did better than most and its hours came out a little ahead. But it is 8 trades.
• Every share that lost sits near half. No word split the winners from the losers by much.
• The 24-hour column is thin, and a mean of a handful of trades can be moved by one of them: "leaning up" averaged +2.07% while 6 of its 10 lost.

Verdict: still not shown. Five looks now, and none has made the case for a gate, so it stays unbuilt. The limits above still hold, and more so: eight days, longs only, in one market's moods.

Credits

• J. Welles Wilder Jr., New Concepts in Technical Trading Systems, Trend Research, 1978: the ADX and its +DI and −DI lines.
• Leonard E. Baum and Ted Petrie, "Statistical Inference for Probabilistic Functions of Finite State Markov Chains", Annals of Mathematical Statistics, 1966: the hidden Markov model.
• James D. Hamilton, "A New Approach to the Economic Analysis of Nonstationary Time Series and the Business Cycle", Econometrica, 1989: switching between hidden states in economics.
• The market words and the look are our own design; the models are described in our page "Models for reading a market, and what they showed on real prices".
• Data: our own paper bots' trades and their hourly market readings, on exchange demo accounts.
• PerpGremlin and Claude, Does the market's mood warn of a bot's losses?, TerminalGremlin, 2026: the tests, the code and the words.