trader-timemachine

The Time Machine sits between the plan and the money. trader-exit has set the entry, the stop and the target. Before anything is sized or sent, this skill asks one question: if today plays out like the stock's own last 60 days, how often does this plan hit the target first, the stop first, or neither? It answers with 10,000 replays and with the real record of the same rule on every past session. It is a simulation built from the stock's own recent bars, not a prediction. Day trading loses most people money, and a good-looking replay does not change that.

${CLAUDE_SKILL_DIR} below is this skill's folder (where this SKILL.md lives).

python3 "${CLAUDE_SKILL_DIR}/timemachine.py" run                    # trader/plan.json, 10,000 replays, seed 19
python3 "${CLAUDE_SKILL_DIR}/timemachine.py" run --write            # also save trader/timemachine.json
python3 "${CLAUDE_SKILL_DIR}/timemachine.py" run --until 15:55      # end each replay at your flat-by time
python3 "${CLAUDE_SKILL_DIR}/timemachine.py" run --json
python3 "${CLAUDE_SKILL_DIR}/timemachine.py" run --symbol TSLA --date 2026-10-19 \
    --entry 364.18 --stop 358.32 --target 370.04 --entry-time "10:15 AM"
python3 "${CLAUDE_SKILL_DIR}/timemachine.py" chart --from           # ASCII fan chart of the saved run

Other flags: --replays N, --seed N (default 19), --window N (default 2 bars), --offline FILE (a saved Yahoo chart JSON, .json or .json.gz), --now "2026-10-19 10:15" (New York time) to pin the clock.

What it reads and writes

How it works

1. The pool. It keeps only FULL sessions before the plan's date: all 78 five-minute bars, 9:30 to 4:00 New York time. Half days, gaps and today are left out. Every bar becomes four ratios to the bar before it: open, high, low and close over the previous close.
2. One replay. Start at the entry price. For each bar left in the day, pick a past bar from the same time of day (plus or minus 2 bars) at random, and scale it to the current price. Bar by bar, it checks the stop first, then the target. A bar that touches both counts as the stop.
3. 10,000 replays give four odds. Target first. Stop first. Neither, which means the trade was still open at --until (default 4:00 PM, the close, or 1:00 PM on an NYSE early-close day). Green, which means target first or above the entry at --until. R is +1 at a 1R target, -1 at the stop, and otherwise the distance from the entry at --until divided by the risk. A gap through the stop still counts as -1R here. A real fill can be worse.
4. The history. The plan's own rule (the same opening range, stop rule and target R) runs again on every one of those past sessions. That is what really happened, not a simulation.
5. The verdict is one line from the numbers. It looks only at the replays that hit one side, and asks how often the target came first. A bigger target needs to win less often to break even: 1 in (1 + R). That is 50 % at 1R, 33 % at 2R, 25 % at 3R. Less than 5 points from break-even: "Close to a coin flip" at 1R, "Close to break-even" at any other R. 5 to 15 points: "Leans toward the target" or "Leans toward the stop". 15 or more: "Strongly toward" one of them. So a 2R plan with target first 24.1 % and stop first 50.5 % is close to break-even (it needs 33 % of the 74.6 % that hit a side, and got 32 %), not a loser.
6. Below the verdict, when it applies: a line when more than 40 % of the replays hit neither (the price at --until decides those), and the average R when its sign goes against the lean. Those can only point the other way because of the replays that hit neither.

The seed. The draws come from Python's random.Random(seed), seed 19 by default. Same plan, same data, same seed: the same numbers, on any machine. Say the seed with every result so the user can repeat it.

5-minute bars only. Yahoo keeps 60 days of 5-min bars but only about 7 days of 1-min bars, too few sessions for a fair pool. It refuses, in one plain line, when there are fewer than 20 full sessions, when the entry bar is not in the data (wrong date, not closed yet, a future day), or when the plan does not make sense (a stop above a buy's entry, an entry off the 5-minute grid). A live in-progress bar is never used.

Steps

1. Make sure trader/plan.json is today's plan from /trader-exit. If there is none, run that first, or pass the plan as flags.
2. Run timemachine.py run --write. Add --until 15:55 when the plan has a flat-by time, so "neither" means still open at the time you would get out.
3. Tell the user, in this order: the four odds, the median and the 5 to 95 range of the price at the close, the average R, the history line, then the verdict line and any line under it exactly as printed. Say the seed.
4. If the verdict is a coin flip or close to break-even, say so plainly. Do not dress it up. If the history and the replays disagree, say that too.
5. Say once that this is built from the last 60 days and is not a prediction. Earnings, news and halts can do things those 60 days never did. Check /trader-news for the stock before going on.
6. If the user still wants the trade, hand off to /trader-budget to size it.

Example: the reel

TSLA, Monday 2026-10-19. The 15-minute opening range is 354.47 to 362.18. The 10:10 bar closes above it: entry 364.18 at 10:15 AM (bar 8), stop 358.32 (the range midpoint, rounded down), target 370.04 (1R, risk 5.86 a share).

10,000 replays of the rest of the session, 10:15 AM to 4:00 PM (69 five-minute bars), seed 19.
  target first   29.4 %
  stop first     25.5 %
  neither        45.1 %   still open at 4:00 PM
  green          51.7 %   target first, or above the entry at 4:00 PM
  Price at 4:00 PM: median 364.36, 90 % of replays between 355.54 and 373.54.
  Average R: +0.034
History: the same rule on each of those 59 real sessions fired 33 times:
13 target, 15 stop, 5 neither. Average R -0.062.
Close to a coin flip: target first 29.4 % vs stop first 25.5 %.
45.1 % of the replays hit neither by 4:00 PM; the price at 4:00 PM decides those.

At 1R the target has to come first in half of the replays that hit one side. It did in 53.6 %. The replays came out close to a coin flip, and the real history was slightly worse than even. That is the honest answer for this setup, and it is what the user should hear before /trader-budget sizes 8 shares.

Rules

• Every number comes from the run. Never round a coin flip up into an edge.
• The verdict line and the lines under it are printed by the tool. Repeat them, do not rewrite them. Never judge a 2R or 3R plan by target first vs stop first alone: the verdict already weighs the target's R.
• Simulated odds are not predictions and not promises. Say it once, plainly.
• Free Yahoo data is unofficial and can lag. Bars are 5-minute bars, checked bar by bar, never tick by tick and never real time.
• Always say the seed, so any run can be repeated.
• One plan per run. A changed entry, stop, target or date is a new run.

What it never does

It never places an order, never sizes a position, never changes the plan, and never calls a result a sure thing. It never prints a P&L or a profit. It never uses a bar that has not closed yet, and never mixes today's bars into the pool.

Safety

• Education and tooling, not financial advice. A PASS is a rules check, not a recommendation.
• Paper trading by default ("paper": true). Going live is the user's own decision, in their own config.
• Nothing in this pack places an order except trader-broker, and only when all three hold: trader-risk passed that exact order today, the user typed an explicit yes for that one order in the chat, and the user clicked Allow once on Claude's own card because the order tool is set to "Needs approval" (Customize > Connectors > tool permissions). Never set an order tool to "Always allow".
• No options, margin, short selling or leverage unless the config allows it and the user confirms.
• Broker keys live in environment variables or the Claude Desktop config, never in the chat.