# Overfit > A public archive of trading strategies that failed our tests, with the receipts. A product of Catalyst Forge, LLC. Tagline: Where Sharpe ratios go to die. Research and education only, not investment advice. We report what failed, not what works. This file is a map for people and for agents. The machine-readable index of every write-up is [https://overfit.bot/graveyard.json](https://overfit.bot/graveyard.json). A slightly richer catalog is [https://overfit.bot/llms-full.txt](https://overfit.bot/llms-full.txt). If you cite a write-up, please link back to it. You may quote short excerpts with a link. Please don't republish the site. ## What this site is, and is not Overfit publishes trading ideas that failed our tests, and shows the work. It is not a signal service, not a system for sale, and not a live trading record. Everything shown is paper research. We report what failed, not what works. A failed test means we could not find it with this data and this rule, not that it can never work. A result labeled "passed its test, no edge" still did not beat simply holding the market on the same days, and has not been traded with real money. ## How to navigate - [Graveyard](https://overfit.bot/graveyard) — every write-up, grouped by cause of death. 17 write-ups as of this build. - [How we test](https://overfit.bot/method) — pre-registration, year splits, trial counts, deflated Sharpe, and the other gates, in plain words. - [Glossary](https://overfit.bot/glossary) — the terms used on the site. - [Passed its test, no edge](https://overfit.bot/survived) — the one idea that cleared the test it was written for but did not beat holding the market on the same days. - [About](https://overfit.bot/about) — publisher, data sources, and corrections. - [Legal](https://overfit.bot/legal) — the full disclaimer. - [RSS](https://overfit.bot/rss.xml) — new write-ups. ## Causes of death - **Overfit** (`overfit`): The winner doesn't survive an honest count of how many things we tried, or it's no better than scrambled data. - **Edge decayed** (`edge-decayed`): It worked in one period and failed a later window we fixed in advance. - **Inconsistent** (`inconsistent`): The sign or result flipped between the windows we fixed in advance, so we don't treat it as a real effect. - **Not tradable** (`not-tradable`): The effect may be real but we couldn't have captured it with realistic trading rules. - **Works but not worth it** (`not-worth-it`): It does something real, but misses its bar or the protection costs more return than we said we'd accept. - **Passed its test, no edge** (`passed-no-edge`): It cleared the pre-registered test it was written for, but did not beat simply holding the market on the same days. It has not been traded with real money. - **Untestable** (`untestable`): A curiosity we would rather mark untestable than invent a p-value for. Each write-up has exactly one primary cause. Causes are explained on [How we test](https://overfit.bot/method); terms are in the [glossary](https://overfit.bot/glossary). ## Write-ups - [1,192 indicator combinations: the winner was no better than luck](https://overfit.bot/posts/grid-1192-combinations): We searched every 1-to-3 indicator combination from 12 signals. The best looked fine until we ran the same search on scrambled data. Cause of death: Overfit. - [144 signal tests, two winners, no edge](https://overfit.bot/posts/ic-screen-144-tests): We screened 24 popular indicators at three horizons. Two passed a false-discovery check. Neither beat what pure chance produces across the whole family. Cause of death: Overfit. - [32 trend variants: the timing was real, picking the winner was not](https://overfit.bot/posts/trend-regime-32-variants): The best trend variant looked superb on 2006–2016. The overfitting checks said it was mostly luck among near-identical cousins. Cause of death: Overfit. - [Indicator families and pairs: best was good, not distinguishable from luck](https://overfit.bot/posts/signal-families-28-books): Momentum plus risk led 28 family and pair combinations. It cut the worst drop in half and still could not clear the overfitting checks. Cause of death: Overfit. - [Machine learning on 24 indicators: no better than shuffled data](https://overfit.bot/posts/walk-forward-models): Regularized regressions and a small tree model, retrained every year, did no better than the same models trained on scrambled answers. Cause of death: Overfit. - [Wait for the stars to align: 24 setups, 0 passed](https://overfit.bot/posts/confluence-24-setups): Sit in T-bills until several conditions line up, then hold for 1, 3 or 6 months. None passed once we counted the tries. Cause of death: Overfit. - [Weekly reversal: great on 2005–2016, nothing on 2017–2021](https://overfit.bot/posts/weekly-reversal): Buy the most oversold ETFs and hold a few weeks. The training years loved it. The pre-registered validation years did not. Cause of death: Edge decayed. - [Yield-curve alarm: right in 2017–2021, almost all cash in 2022–2023](https://overfit.bot/posts/yield-curve-gate): Go to cash when the yield curve inverts. It helped in one window, then sat out most of the next. Cause of death: Edge decayed. - [Moon phases: the sign flipped; not confirmed](https://overfit.bot/posts/lunar-cycle): A sign test on lunar-phase returns. It failed the 2017–2021 window and parked the whole astro-timing family. Cause of death: Inconsistent. - [Pre-Fed-announcement drift: real, but it happens overnight](https://overfit.bot/posts/pre-fomc-drift): The premium before Fed meetings is documented. In our data the gain arrived overnight, before a next-morning fill could catch it. Cause of death: Not tradable. - [A take-or-skip judge on top of our setups: it took nearly everything](https://overfit.bot/posts/take-or-skip-gate): We asked a decision gate to approve or reject signals from our earlier setups. It either approved nearly every one or its picks were no better than its skips. Cause of death: Works but not worth it. - [The plain 200-day rule: smaller drops, not enough to pass](https://overfit.bot/posts/sma200-overlay): One simple, untuned rule written down in advance. It passed one validation window and failed the other. Cause of death: Works but not worth it. - [Turn-of-month plus Fed days: more trades, same disappointment](https://overfit.bot/posts/tom-plus-fomc): Combining the two calendar effects did not fix either. Cause of death: Works but not worth it. - [Turn-of-the-month: invested 12% of the time, earned 4% a year](https://overfit.bot/posts/turn-of-month): The turn of the month has a documented return premium. As a standalone rule it left money in cash almost all the time. Cause of death: Works but not worth it. - [Volatility-targeting plain SPY, VTI and VUG: close, not enough](https://overfit.bot/posts/vol-target-core-etfs): Scale an equal-weight mix of three broad funds to a 10% volatility target. It fell short of the pre-set Sharpe bar: a 0.035 shortfall where 0.02 was allowed. Cause of death: Works but not worth it. - [The drawdown brake: passed its test, no edge](https://overfit.bot/posts/v-overlay-drawdown-brake): It passed the pre-registered test vs the plain 200-day rule, including 6 of 7 new markets. It did not beat holding SPY on the same days, and it has not been traded with real money. Cause of death: Passed its test, no edge. - [The Benner cycle (1875): beautiful chart, untestable](https://overfit.bot/posts/benner-cycle): A famous chart of "good times" and "hard times" years. We have about 21 years of daily data, so we cannot test it statistically. Cause of death: Untestable. ## For agents - Machine-readable index: [https://overfit.bot/graveyard.json](https://overfit.bot/graveyard.json) (one record per write-up). - Published receipts as JSON: [https://overfit.bot/graveyard-receipts.json](https://overfit.bot/graveyard-receipts.json). This is only what the write-ups already show (tables, windows, cost checks, same-days SPY rows) plus short excerpts of each page's idea, method and reason it failed, with the link back. It has no price series and no per-trade data. - A free, read-only MCP server (Streamable HTTP, no authentication, rate limited, beta) is at https://mcp.overfit.bot/mcp. It answers from the same published data as this site, fetched live, so it cannot disagree with the pages. Tools: - `check_strategy_family`: Is this kind of idea already in the graveyard? Matches by idea, not by parameters. - `get_burial_receipt`: The published receipts for one write-up: cause of death, windows, cost checks, same-days SPY rows. - `get_test_protocol`: How we test: windows, pre-registration, the bars, and the limits. - `overfit_checklist`: Deterministic trial-count and deflated-Sharpe arithmetic for your own result. Stateless. It reports what failed, not what works; every answer carries a disclaimer and a link to the write-up. It cannot tell you that a strategy will work. ## Data and licensing Based on public daily price data from Tiingo, and from FRED, the Federal Reserve Board, and Coinbase where a write-up says so. We publish derived statistics only and do not republish the underlying price series. © 2026 Catalyst Forge, LLC. You may quote short excerpts with a link back. Please don’t republish the site. Tiingo, FRED, the Federal Reserve and Coinbase are not affiliated with Overfit and do not endorse it. ## Corrections If you find a mistake or a data-rights concern, see [corrections](https://overfit.bot/about#corrections). We note corrections on the affected page.