Desk field guide
Each entry says what the thing is, how it's worked out, and why an agent would use it. Where it helps there's a small example with made-up round numbers.
1. The basics everything else is built on
One slice of time boiled down to five numbers: the price it opened at, its high, its low, where it closed, and the volume traded. A "1-minute bar" covers one minute and a "1 h bar" covers one hour. Every indicator below is arithmetic on a run of bars.
The playground feeds each idea Coinbase's 1-minute Bitcoin bars. Only finished bars go in, so an idea can never peek at a minute that hasn't ended.
One hundredth of a percent. 100 bp = 1%. A move of 25 bp on a $100 price is 25 cents.
Agents use basis points so a move on Bitcoin and a move on oil can be compared on the same ruler.
The bid is the best price someone will pay right now and the ask is the best price someone will sell at. Together they're the "touch", and the gap between them is the spread. Buying now means paying the ask, and selling now means taking the bid.
The playground fills its paper trades at the touch, not at the last trade price, because that's what you'd actually get. Crossing the spread costs money on every round trip.
Kalshi charges a fee on each side of a trade (in and out). Perps also charge or pay funding at set times, a small payment between longs and shorts that keeps the perp's price near the real asset's.
A trade that wins 3 bp of price but pays 4 bp in fees is a loser. That's why the new take-profit is measured after fees.
Pretend trades written down at real prices, with no money and no orders sent. Everything in the playground is paper.
2. Exits: how a trade gets closed
"Close it after N minutes, no matter what." Every idea has to set one. It's the safety cap, so no paper trade can sit forever.
"Close it once it's up this much." Today the playground measures this as a raw price move. Coming: a take-profit on profit after fees and funding, which is what actually lands in the account.
"Close it if it's down this much." Today it's a fixed number of basis points. Coming: a stop set in ATRs (next section), so it widens on wild days and tightens on calm ones.
The stop distance is "1R". A target at "1.5R" is one and a half times as far as the stop. The Setups page aims for 1.5R: risk $10 to make $15.
A stop that follows the price as it moves your way and never moves back. The Setups "runner" plan sells half at the target, then trails the rest one ATR behind the best price reached.
3. ATR: average true range
In one line: ATR is how far the price typically moves in one bar. It's a "normal wiggle" ruler for the market right now.
Step 1: the true range of one bar
The plain range is high minus low. "True" range also counts a gap from the previous bar's close, so a jump between bars isn't missed. It's the biggest of three numbers:
|high − previous close|
|low − previous close|
Example: the last hour closed at 100. This hour's high is 103 and its low is 101. High − low is only 2, but the price jumped from 100 and reached 103, so the true range is 3.
Step 2: average it (Wilder's way)
Take the plain average of the first 14 true ranges. After that, each new bar moves the average only a little: new ATR = (old ATR × 13 + today's true range) ÷ 14. One wild bar nudges ATR up without yanking it. This smoothing is named after J. Welles Wilder, who invented ATR in 1978.
Step 3: use it for a stop
Put the stop k ATRs from your entry. With ATR at 2.50 and k = 1, a long bought at 100 has its stop at 97.50. On a calm day ATR might be 1.00, so the stop sits at 99. The stop fits the market's mood instead of being one fixed number that's too tight on wild days (you get shaken out by noise) and too loose on calm ones (you lose more than you needed to).
The Setups page uses one 1-hour ATR, smoothed over 14 bars. The agents will be able to choose their own: how many ATRs away, how many bars, what bar size (1 minute to 4 hours), smoothed or plain average, and fixed or trailing. More choices aren't free: every setting an agent tries counts against it (section 6).
The same formula on different bars gives different numbers. The playground uses Coinbase's prices and the Setups page uses Kalshi's, so their ATRs will be close but not equal. Some desk pages (Time Machine replay, Move Atlas, the live heat map) use a plain 24-hour average with no smoothing, so their "ATR" is different again.
4. The playground's three ideas so far
Each idea is one number worked out from the last day of 1-minute bars. When the number crosses a line, the idea buys one paper contract, then sells after its hold time.
The number: over the last 15 minutes, count each red minute's volume (closed below its open) as minus and each green minute's as plus, then average it. Divide by a normal minute's volume over the last day. −2 means sellers have been dumping at twice normal volume, every minute, for 15 minutes.
The bet: at −2 or below, buy, betting the selling has worn itself out and price bounces. Sell 30 minutes later.
This is a crude order flow measure: who's pushing harder, buyers or sellers.
The number: how far Bitcoin has moved over the last 24 hours, divided by how far random minute-to-minute wandering would normally carry it in a day. +1.5 means it rose one and a half times further than normal noise.
The bet: at +1.5 or above, buy, betting a strong day keeps going (momentum). Sell 2 hours later.
The number: how jumpy the last 30 minutes have been compared with the whole day. 2 means twice as jumpy. It says nothing about direction.
The bet: at 2 or above, buy, betting that sudden activity tends to resolve upward. Sell 4 hours later.
Volatility is how big the moves are, not which way they go. ATR is one way to measure it, and this ratio is another.
None of the three uses a take-profit or stop yet, only hold time. The first three paper trades (Nov 4) lost about 8 cents in total, one winner and two losers. That's far too few to judge anything.
5. The Charts page and the Setups page
An exponential moving average is the average of recent closes, weighted so newer bars count more. The 20-bar EMA reacts fast and the 50-bar EMA reacts slowly. Price above both, with the 20 above the 50, is the classic picture of an uptrend.
Volume-weighted average price: the average price paid since a starting point (the session open, the day, the last 24 hours), with busy moments counting more. Big traders treat it as "fair price today". Above VWAP, buyers have the upper hand on the day. Below it, sellers do.
A line through the last 20 closes' average, with bands 2 standard deviations above and below. Price outside a band is unusually far from its recent average. Narrow bands mean a quiet market, which often comes before a burst.
The Charts page's one-word summary for each timeframe. Up = price above the 20 EMA, the 20 above the 50, and price above VWAP. Down = all three the other way. Mixed = anything else.
The same market viewed as 1-minute, 5-minute, 15-minute, 1-hour, 4-hour, daily, weekly and monthly bars. The fast rows show timing and the slow rows show the bigger trend you're trading with or against.
From Kalshi's own trade tape: contracts bought by aggressive buyers minus contracts sold by aggressive sellers, over the last 5 minutes. It shows who is pushing right now.
A history of past "runs": how long up or down moves usually last and how far they go, by market and time of day. Setups uses it to warn when you'd be joining a move that's already old.
| Line | Plain question |
|---|---|
| 1 | Am I chasing a move that's already hours old? |
| 2 | Is the 1-hour lean against me? |
| 3 | Is price stretched: several ATRs from its 24 h VWAP and outside a Bollinger band? |
| 4 | Does a 1-ATR stop fit my dollar limit and stay clear of liquidation? |
| 5 | Am I buying into a dip (entering against the last 5 minutes' push), not into a spike? |
| 6 | Did I just close a winner? (A reminder: overconfidence after wins.) |
| 7 | Is anything scheduled (news, funding, a market close) before I'd be out? |
Passing lines add up to a grade of A, B, C or "no setup". It's a decision aid, not a proven edge: the agents' own tests found that EMA/VWAP timing didn't beat a plain trend rule. The page records every grade going forward so we'll learn whether A grades really do better.
6. How the agents keep from fooling themselves
The danger in trading research: try enough ideas on old data and some will look brilliant by luck alone. These checks are there to catch that.
Ideas are hunted for on one stretch of history (discovery) and then judged on data they've never seen (holdout, or a forward paper run). If an idea only works on the data it was found on, it was luck.
Before a test runs, the idea's exact rules are written down and fingerprinted (a "hash"). Changing even one number changes the fingerprint, so nobody can quietly tweak the rules after seeing results. A change means a new version, with a note saying what changed and why.
Every setting an agent tried while hunting is counted. Try 470 settings and the best one will look good just by chance. The count goes on the card, and it's why giving agents more exit choices doesn't make passing easier.
How many "noise units" the average result sits away from zero. Around 2 or more starts to look like more than luck for a single test. The strong day trend idea scored 1.79 in discovery: interesting, not proven.
Drop the five best trades and look again. The strong day trend idea went from +35.8 bp a trade to +3.9 bp, so most of its result came from a handful of big days. That's a warning sign worth knowing about.
When several ideas are tested at once, each one has to clear a stricter bar. This uses a standard method called Benjamini–Hochberg. With 3 ideas live, each needs p ≤ 0.033 instead of the usual 0.05. The card shows the bar each idea has to clear.
An idea has to beat doing something dumb, like "always be long". If always-long made the same, the idea added nothing.
Kronos is an open AI model that forecasts prices from bars. The agents tested it on 18 market-and-horizon combinations. It lost to "assume no change" in all 18, so it's kept as background context only, not as a signal. Agents can still ask to try tools like this through the request list.
7. Coming next
- An exit menu for every idea: take-profit after fees, ATR stops with the agent's own settings, and the hold-time cap kept.
- A tool-request list: an agent files "I need X, here's why, for this idea". You and the lead decide what gets built.
- Every new tool comes with its plain-words entry, added to this guide.