<- all tokdocs
2026.10.23@buildwithai.club

The video calls 29.4 percent "the historical odds"; the repo's own README labels that number a 10,000-path simulation and puts the real history at 13 of 33

Finance
Watch on TikTok

View on TikTok ->

A 39.44-second vertical clip, 720x1280, in a QuickTime/MP4 container (ffprobe format_name=mov,mp4,m4a,3gp,3g2,mj2), HEVC Main profile level 120 at pix_fmt=yuv420p with r_frame_rate=25/1 and avg_frame_rate=24475/986 (24.82 fps) across 979 video frames, video bitrate 130,175 bps, audio AAC HE-AACv2 at 44,100 Hz stereo and 64,301 bps across 853 frames with a 39.427-second audio duration, overall container bitrate 199,988 bps, filesize 985,941 bytes, pulled as format id bytevc1_720p_199988-1; posted at epoch 1792763653, which is 2026-10-23T13:54:13 UTC, by uploader buildwithai.club (uploader_id 7597027599110046731, channel nickname "Build With AI"), audio credited to track "original sound" by artist "Build With AI", with a 581-character description; at capture time (metadata.json mtime 2026-10-26 00:39:54) the counters read 209,300 views, 11,500 likes, 174 comments, 2,223 reposts and 11,215 saves. I read 19 of the 20 extracted frames (frame_001 through frame_014 and frame_016 through frame_020) and re-extracted the 7-to-11-second window at fps=4 because the 2-second sampling visibly skipped a card-reveal animation; from that dense pass I read d_006, d_008 and d_010 at roughly 8.25 s, 8.75 s and 9.25 s. The transcript is 12 Whisper segments and 146 words. frame_001 (0 s) is a split screen: the top half is a cream motion-graphic panel with a black pill reading "✦ CLAUDE", a white card headed "Your trade" beside a tan-outlined "EXAMPLE" tag, a dark "Buy" button, a candlestick series with a black "Entry" marker and a dashed horizontal line, and a pixel-art desk with a red character; the bottom half is a man in a navy shirt and clear-framed glasses in a dark room, with the burned-in word "stock". frame_002 (2 s) adds "10,000 replays" in the top right, a padlock over a greyed-out "Buy", a green "Target" line, a red "Stop" line, and a dense spaghetti of thousands of green, blue and red simulated paths. frame_003 (4 s) fills the screen with a cream card showing a red folder icon and the text "day-trader-skill" with a blinking cursor. frame_004 (6 s) adds a black "11 SKILLS" pill, a tan "FREE" pill, a red rubber stamp reading "OPEN SOURCE · MIT", and eleven blank placeholder cards in a 4/4/3 grid. frame_005 (8 s) catches the cards mid-reveal with "Chart Scanner", "News Monitor", "Entry Timer", "Exit Planner", "Time Machine", "Risk Manager", "Budget Manager" and "Order Executor" legible and the bottom row still animating; only the dense frame d_010 (9.25 s) shows all eleven labels at once, adding "Price Alerts", "Earnings Analyst" and "Pelosi Tracker". frames 006 through 012 walk a five-stage nav strip of pills reading "Scanner", "News", "Entry", "Exit", "Time Machine" with one highlighted red at a time: a grid of roughly thirty sparkline cards tagged "THOUSANDS OF CHARTS", then three result rows reading "Stock A / Opening-range breakout", "Stock B / Flag" and "Stock C / Oversold bounce", then "BREAKING" and "EARNINGS" news cards with redacted grey bars, then an eye icon and "Waiting", then a gold "Triggered" badge, then a green "Take profit" band labelled "1× risk" over a "Risk" bracket. frames 013 through 017 are a full-bleed Time Machine sequence: an isometric "YOUR TRADE" card, a counter animating through "2,800 replays of the rest of the day" to "10,000 replays of the rest of the day", a tan-outlined tag reading "EXAMPLE RUN · FROM THE README", and three isometric stacks labelled "29.4% Hit target" and "25.5% Hit stop". frame_019 (36 s) returns to the split screen with a fan chart between "Take profit" and "Stop-loss" lines and a row labelled "SIMULATED ODDS" reading 29.4%, 25.5% and 45.1%, beside a "$1" coin, a padlocked "Buy" and the words "Before any money moves"; frame_020 (38 s) is the talking head alone under the caption "a guarantee." No terminal, no Claude Code session, no shell command, no GitHub page and no repository URL appears in any frame I read; every screen is a custom motion-graphic mockup, and the honest tell is the "EXAMPLE RUN · FROM THE README" tag that concedes the headline percentages were copied out of a README rather than produced by a run on camera. The spoken argument is that somebody released eleven free, open-source Claude skills for day trading, that the standout is a "time machine" which replays the rest of the session 10,000 times from historical price data, and that this shows you "the historical odds" before you risk money. The transcript ends cleanly: the last segment runs 37.34 s to 39.20 s against a 39.427-second audio stream, so nothing is timestamped past the end of the audio and no segment is a Whisper hallucination; Whisper also rendered "Claude" correctly rather than mangling it to "cloud", which the "✦ CLAUDE" pill in frame_001 confirms.

The repository is real, the MIT stamp is accurate, and "just" means three days

At 4.74 s the video says "Someone just open-sourced 11 free day trading skills for Claude." The GitHub API for Jakeschincariol/day-trader-skill returns created_at of 2026-10-20T14:17:54Z and a single commit dated 2026-10-20T14:17:50Z with the message "The Trader skill v2: 11 day-trading skills". The video's own post timestamp is 2026-10-23T13:54:13 UTC, so "just" covers two days and roughly 23.6 hours. The same API call returns license.spdx_id of MIT, which matches the red "OPEN SOURCE · MIT" stamp in frame_004, and fork: false, so this account is the origin rather than a copy.

The eleven cards map one to one onto the eleven top-level skill folders in the repository, with display names that differ from the folder names: trader-scanner is "Chart Scanner", trader-news is "News Monitor", trader-entry is "Entry Timer", trader-exit is "Exit Planner", trader-timemachine is "Time Machine", trader-risk is "Risk Manager", trader-budget is "Budget Manager", trader-broker is "Order Executor", trader-alerts is "Price Alerts", trader-analyst is "Earnings Analyst" and trader-pelosi is "Pelosi Tracker". The two cards the 2-second frame sampling dropped, "Earnings Analyst" and "Pelosi Tracker", are the two that would tell a viewer this pack also reads SEC filings and House Clerk disclosure data, which is a different proposition from chart scanning.

One practical gap: the video never names the author and never shows a URL. Searching the repository name surfaces a wall of byte-identical forks under accounts such as CyberGanna, cabrio5027-max, kaichaoang, madusankapremaratne, camilacz613-prog and schwarzemarkusgmxdemsms-max, all carrying the same description string. The GitHub API reported 1,845 stars, 390 forks and 32 watchers on the original when I checked on 2026-10-26. A viewer following this video has no way, from the video alone, to tell the original from a fork.

The 10,000 replays bound the simulator's own sampling error, not the trade's risk

At 24.58 s the video says "It takes your trade and simulates the rest of the trading day 10,000 times using historical price data." As a description of the code, that is correct. In trader-timemachine/timemachine.py the function signature is simulate(pool, entry, stop, target, k, last, replays=10000, seed=19, window=2, side="buy"), and its docstring reads "Monte Carlo replays of bars k+1..last." The module docstring spells out the draw: "For each bar i it picks a slot j = i + randint(-window, window), kept in 1..77, then one past bar from pool[j] (random.choice), and scales it to the current price." build_pool converts every past session's bar into (o, h, l, c) divided by the previous bar's close. That is a resampling-with-replacement scheme, the bootstrap of Efron (1979), applied to intraday candle ratios.

The problem is what the number 10,000 actually buys. It controls Monte Carlo sampling error and nothing else. At the reported proportion of 0.294, the binomial standard error across 10,000 independent replays is sqrt(0.294 × 0.706 / 10000) = 0.456 percentage points, so the 95 percent half-width is 0.89 percentage points. Running 1,000 replays instead would give a half-width of 2.82 points. Going from 1,000 to 10,000 therefore tightens the estimate by about two percentage points and leaves every other source of error untouched, because all 10,000 paths are drawn from the same fixed pool of 59 past sessions.

The video's opening line at 0.00 s, "Claude can now test your stock trade 10,000 times before you risk a single dollar," reads as though the trade is what gets tested 10,000 times. What gets sampled 10,000 times is the resampler. The repository is clearer than the video on this point: its fine print reads "The Time Machine is a simulation, not a prediction. It is built from the stock's own last 60 days of 5-minute bars. The future can do things those 60 days never did: an earnings report, a halt, a headline. Treat its odds as a reality check, never as a promise." The video compresses all of that into one clause at 37.34 s, "Of course, it's a simulation, not a guarantee."

Drawing each bar independently removes the serial dependence an intraday breakout rule depends on

The sampler picks one historical bar per slot with random.choice, independently for every bar in the path. The ±2 slot window preserves the time-of-day shape of volatility, so a 10:15 bar is drawn from a pool of bars that occurred near 10:15. It does not preserve dependence within a path. In a generated session, a large up bar at 11:00 carries no information about the bar at 11:05.

That matters because the i.i.d. bootstrap is not valid for dependent data, which is exactly why block methods exist: the moving blocks bootstrap of Künsch (1989) and the stationary bootstrap of Politis and Romano, Journal of the American Statistical Association 89, pages 1303 to 1313 (1994), both cited as such in Politis and White's 2004 paper on automatic block-length selection. Volatility clustering, where large moves follow large moves, is one of the standard empirical properties of asset returns catalogued by Cont in Quantitative Finance volume 1, issue 2, pages 223 to 236 (2001). Independent per-bar draws reproduce the marginal distribution of candle ratios and discard the clustering.

The direction of the resulting bias is visible in the repository's own numbers. A path that never strings together consecutive moves in one direction is less likely to travel far enough to touch either the target or the stop, so outcomes pile up in the "neither" bucket. The simulation reported "still open at the close 45.1 %". The same rule run on the 59 real sessions produced 5 "neither" out of 33 trades, which is 15.2 percent. That is a 30-point gap in the one bucket the method's independence assumption would be expected to inflate.

The repository's two methods disagree on this exact trade, and the video shows only one of them

The README walks through a worked example it labels "the trade in the reel: Tesla, Monday October 19 2026, real data." Step 5 reads: "/trader-timemachine replayed the rest of the day 10,000 times, each 5-minute candle drawn from TSLA's own candles at that time of day over the previous 59 sessions (seed 19, so anyone can repeat it): target first 29.4 %, stop first 25.5 %, still open at the close 45.1 %. It also re-ran the same rule on those 59 real sessions: it fired 33 times, 13 hit the target, 15 the stop, 5 neither. Its verdict: close to a coin flip."

Those are the two sides of the comparison. The simulated triple is 29.4 / 25.5 / 45.1. The historical triple is 13/33, 15/33 and 5/33, which is 39.4 / 45.5 / 15.2. They disagree by ten points on target, twenty points on stop and thirty points on "neither". frames 016 through 019 of the video display only the simulated triple. The historical triple appears in no frame I read.

Neither side is precise enough to settle the question, and the mismatch in precision is the trap. The 95 percent Wilson interval on 13 of 33 runs from 24.7 percent to 56.3 percent; on 15 of 33 it runs from 29.8 percent to 62.0 percent; on 5 of 33 it runs from 6.7 percent to 30.9 percent. A 33-trade sample cannot distinguish a good setup from a bad one. The simulated figure carries a 0.89-point half-width and therefore looks far sharper, but that sharpness describes the simulator's internal consistency, not the market.

The specific correction is to the line at 33.98 s: "So you get to see the historical odds before putting your money on the line." The numbers the video puts on screen are not historical odds. The video's own graphic at frame_019 labels the row "SIMULATED ODDS", and the README labels them as the output of 10,000 replays. The historical odds from the same README are 39.4 percent to target, not 29.4 percent. The voiceover and the video's own caption contradict each other.

No run appears on camera, and no configuration count appears anywhere

The spec question for a backtesting pitch is whether the tool is shown doing the thing. It is not. Across frames 001 to 020 there is no terminal, no python3 trader-timemachine/timemachine.py run, no Claude Code transcript, no file tree and no browser. Every surface is a bespoke motion graphic, and three separate screens carry an "EXAMPLE" tag (frames 001, 009, 010, 018, 019) while frames 016 and 017 carry "EXAMPLE RUN · FROM THE README". The tag is to the creator's credit, because it states plainly that the 29.4 and 25.5 figures were not generated for this video. It also means the video contains zero independent evidence that the tool runs.

A second, smaller overstatement sits at 9.94 s: "One skill scans thousands of stock charts looking for trading opportunities", reinforced by the burned-in card "THOUSANDS OF CHARTS" in frame_006. The README's worked example reports a scan of 231 charts, of which 42 had closed above their 15-minute opening range. The scanner can be pointed at a larger universe, since it reads Nasdaq Trader symbol lists, but the only run the repository documents scanned 231.

The deeper issue is one neither the video nor the repository addresses. Bailey, Borwein, López de Prado and Zhu, in "The Probability of Backtest Overfitting" (revised February 2015), write that backtests "almost never declare the number of trials involved in a discovery. Because those researchers have most likely not controlled for the number of trials, it is highly probable that their findings constitute false positives." The documented example fixes a specific configuration: a 15-minute opening range, a stop at the range midpoint, a target at one times risk, flat by 3:55 PM. Nothing in the README, the SKILL.md or the video says how many configurations were tried before that one, and the Time Machine reports no statistic that would let a user assess it. The tool's verdict function compares the target's share of decided outcomes against the break-even share implied by the target R, which is a useful sanity check, and it is not a control for selection over configurations.

To the repository's credit, its fine print discloses the limits the video omits: "Most people who day trade lose money", the Yahoo Finance chart endpoint is "not an official API: it can lag, change or rate-limit you", only 5-minute bars are used because Yahoo keeps roughly 7 days of 1-minute bars, the run refuses below 20 full sessions, a bar that touches both stop and target is counted as a stop, and a gap through the stop is scored as exactly -1R while "a real fill can be worse." The 39-second video carries none of that.

Key Takeaways

  • Verified: The repository exists, is MIT licensed, and contains exactly eleven skill folders matching the eleven cards on screen. GitHub's API reports created_at 2026-10-20T14:17:54Z against the video's 2026-10-23T13:54:13 UTC post time, so "someone just open-sourced" is accurate to within three days.
  • Verified: The Time Machine does default to 10,000 replays drawn from the stock's own prior 5-minute bars. simulate() carries the signature replays=10000, seed=19, window=2, and the fixed seed makes the README's figures reproducible as claimed.
  • Correction: The video says at 33.98 s "you get to see the historical odds." The on-screen 29.4 / 25.5 / 45.1 are simulated odds, labelled "SIMULATED ODDS" in the video's own frame_019 and described in the README as the output of 10,000 replays. The README's actual historical replay of the same rule gives 13 of 33 to target (39.4 percent), 15 of 33 to stop (45.5 percent) and 5 of 33 neither (15.2 percent).
  • Correction: The video never shows the historical replay, which is the half of the repository's output that reports what happened rather than what was simulated. The two halves disagree by 10, 20 and 30 percentage points on the same trade.
  • Clarification: 10,000 replays bounds Monte Carlo sampling error to a 95 percent half-width of 0.89 points at p = 0.294. It does not address the uncertainty from the 59-session pool. The historical comparison the simulation is checked against is 33 trades, whose 95 percent Wilson interval for 13 of 33 runs from 24.7 percent to 56.3 percent.
  • Clarification: The sampler draws each bar independently via random.choice, which discards serial dependence. The i.i.d. bootstrap is invalid for dependent data, which is why Künsch (1989) and Politis and Romano (1994) developed block methods, and volatility clustering is a documented property of returns (Cont 2001). The 45.1 percent "neither" bucket against a historical 15.2 percent is consistent with that bias.
  • Partial correction: "Scans thousands of stock charts" at 9.94 s is a capability claim, not a documented run. The README's worked example scanned 231 charts and found 42 above their opening range.
  • Context: No frame I read shows a terminal, a Claude Code session, a repository page or a URL. The 30-second card reading "EXAMPLE RUN · FROM THE README" states outright that the headline percentages were copied from documentation rather than generated on camera.
  • Context: No configuration count is reported anywhere in the tool or the video. Bailey, Borwein, López de Prado and Zhu argue that an undeclared trial count makes overfitting impossible to assess.
  • Context: The repository discloses that most day traders lose money, that the Yahoo chart endpoint is unofficial and can lag, that a bar touching both stop and target counts as a stop, and that a gap through the stop can fill worse than -1R. The video carries one hedge, at 37.34 s, and nothing else.
  • Context: The 2-second frame sampling dropped two of eleven skill cards. Only a 4 fps re-extraction of the 7-to-11-second window revealed "Earnings Analyst" and "Pelosi Tracker", the two skills that read SEC filings and congressional disclosures rather than price charts.

Resources

Published October 23, 2026. Writeup generated from a favorited TikTok.