Stenly Tenev

I test market claims against SEC and FINRA data and publish the ones that fail. Including my own. Numbers from primary sources, never headlines.

[Stenly Tenev @StenlyTenev]

Everyone watches price. Almost nobody watches who is trading.

FINRA publishes weekly off-exchange trade counts per ticker. Free. Retail order flow runs through wholesalers, so this is a clean proxy for small investor activity.

I measured 552 events where a stock's retail trade count doubled against its own 12-month baseline.

Signal months: prior 12m +6.4%, next 12m +25.6%. Total 32.0%
Control months: prior 12m +13.5%, next 12m +18.1%. Total 31.6%

The totals are the same. The signal does not predict more return. It predicts when the return arrives: 57% of it lands after a control month, 80% after a signal month.

So the forward edge is +7.5 points, not +25.6%.

Then I tried to kill it. Within-company, each company's signal months against its own quiet months: positive in 76% of companies, median +10.5pp, p < 0.001. It survived.

But this died: it does NOT predict which company doubles.
23% of signal companies doubled. 23% of all companies doubled. Identical.

It tells you when. Not which.

[Stenly Tenev @StenlyTenev]

A coding agent forgets everything between sessions. Not some of it, all of it: what was decided, what was ruled out, what was tried and failed.

"Write it down" fails exactly in the long, loaded session where it matters most. So I made the writing structural instead of a good intention.

Baton is a Claude Code plugin. Each task is a folder with a logbook. Hooks, run by the harness rather than by the model's goodwill:
- at session start, the agent sees which tasks were touched last and where each one stopped;
- at the end of a turn, if a task folder holds work newer than its logbook, the turn goes back to the agent with the name of that folder;
- after context compaction, the agent writes the logbook first.

What it does not do: the hooks check that the logbook was written, not that it is right. A wrong entry passes.

Free, MIT. In Anthropic's plugin directory, or:
claude plugin marketplace add StanimirTenev/baton
claude plugin install baton@baton

[Stenly Tenev @StenlyTenev]

A coding agent forgets everything between sessions. Not some of it, all of it: what was decided, what was ruled out, what was tried and failed.

"Write it down" fails exactly in the long, loaded session where it matters most. So I made the writing structural instead of a good intention.

Baton is a Claude Code plugin. Each task is a folder with a logbook. Hooks, run by the harness rather than by the model's goodwill:
- at session start, the agent sees which tasks were touched last and where each one stopped;
- at the end of a turn, if a task folder holds work newer than its logbook, the turn goes back to the agent with the name of that folder;
- after context compaction, the agent writes the logbook first.

What it does not do: the hooks check that the logbook was written, not that it is right. A wrong entry passes.

Free, MIT. In Anthropic's plugin directory, or:
claude plugin marketplace add StanimirTenev/baton
claude plugin install baton@baton

[Stenly Tenev @StenlyTenev]

A coding agent forgets everything between sessions. Not some of it, all of it: what was decided, what was ruled out, what was tried and failed.

"Write it down" fails exactly in the long, loaded session where it matters most. So I made the writing structural instead of a good intention.

Baton is a Claude Code plugin. Each task is a folder with a logbook. Hooks, run by the harness rather than by the model's goodwill:
- at session start, the agent sees which tasks were touched last and where each one stopped;
- at the end of a turn, if a task folder holds work newer than its logbook, the turn goes back to the agent with the name of that folder;
- after context compaction, the agent writes the logbook first.

What it does not do: the hooks check that the logbook was written, not that it is right. A wrong entry passes.

Free, MIT. In Anthropic's plugin directory, or:
claude plugin marketplace add StanimirTenev/baton
claude plugin install baton@baton

[Stenly Tenev @StenlyTenev]

A coding agent forgets everything between sessions. Not some of it, all of it: what was decided, what was ruled out, what was tried and failed.

"Write it down" fails exactly in the long, loaded session where it matters most. So I made the writing structural instead of a good intention.

Baton is a Claude Code plugin. Each task is a folder with a logbook. Hooks, run by the harness rather than by the model's goodwill:
- at session start, the agent sees which tasks were touched last and where each one stopped;
- at the end of a turn, if a task folder holds work newer than its logbook, the turn goes back to the agent with the name of that folder;
- after context compaction, the agent writes the logbook first.

What it does not do: the hooks check that the logbook was written, not that it is right. A wrong entry passes.

Free, MIT. In Anthropic's plugin directory, or:
claude plugin marketplace add StanimirTenev/baton
claude plugin install baton@baton

[Stenly Tenev @StenlyTenev]

A coding agent forgets everything between sessions. Not some of it, all of it: what was decided, what was ruled out, what was tried and failed.

"Write it down" fails exactly in the long, loaded session where it matters most. So I made the writing structural instead of a good intention.

Baton is a Claude Code plugin. Each task is a folder with a logbook. Hooks, run by the harness rather than by the model's goodwill:
- at session start, the agent sees which tasks were touched last and where each one stopped;
- at the end of a turn, if a task folder holds work newer than its logbook, the turn goes back to the agent with the name of that folder;
- after context compaction, the agent writes the logbook first.

What it does not do: the hooks check that the logbook was written, not that it is right. A wrong entry passes.

Free, MIT. In Anthropic's plugin directory, or:
claude plugin marketplace add StanimirTenev/baton
claude plugin install baton@baton

[Stenly Tenev @StenlyTenev]

A coding agent forgets everything between sessions. Not some of it, all of it: what was decided, what was ruled out, what was tried and failed.

"Write it down" fails exactly in the long, loaded session where it matters most. So I made the writing structural instead of a good intention.

Baton is a Claude Code plugin. Each task is a folder with a logbook. Hooks, run by the harness rather than by the model's goodwill:
- at session start, the agent sees which tasks were touched last and where each one stopped;
- at the end of a turn, if a task folder holds work newer than its logbook, the turn goes back to the agent with the name of that folder;
- after context compaction, the agent writes the logbook first.

What it does not do: the hooks check that the logbook was written, not that it is right. A wrong entry passes.

Free, MIT. In Anthropic's plugin directory, or:
claude plugin marketplace add StanimirTenev/baton
claude plugin install baton@baton

[Stenly Tenev @StenlyTenev]

A coding agent forgets everything between sessions. Not some of it, all of it: what was decided, what was ruled out, what was tried and failed.

"Write it down" fails exactly in the long, loaded session where it matters most. So I made the writing structural instead of a good intention.

Baton is a Claude Code plugin. Each task is a folder with a logbook. Hooks, run by the harness rather than by the model's goodwill:
- at session start, the agent sees which tasks were touched last and where each one stopped;
- at the end of a turn, if a task folder holds work newer than its logbook, the turn goes back to the agent with the name of that folder;
- after context compaction, the agent writes the logbook first.

What it does not do: the hooks check that the logbook was written, not that it is right. A wrong entry passes.

Free, MIT. In Anthropic's plugin directory, or:
claude plugin marketplace add StanimirTenev/baton
claude plugin install baton@baton

[Stenly Tenev @StenlyTenev]

@SouthernValue95 Broadcom's actual gross margin is already 69.5%, with a 48.6% operating margin last quarter. Mid-60s is the conservative end of that.

[Stenly Tenev @StenlyTenev]

@amitisinvesting The earnings side backs that. Home Depot reported this morning: $47.86B of sales, +5.7%, comparable sales +1.7% against +0.7% expected.

[Stenly Tenev @StenlyTenev]

@AIStockSavvy Duolingo's revenue growth halved over four quarters: 41.5%, 41.1%, 26.5%, 18.3%. The June operating margin slipped too, 13.2% to 11.4%.

[Stenly Tenev @StenlyTenev]

@pequityresearch Micron's last reported quarter already ran an 84.6% gross margin, up from 37.7% a year earlier. The 90% call is six points from here.

[Stenly Tenev @StenlyTenev]

@babyfolio Amtech's June quarter across three years: sales $30.7M to $22.4M, gross profit $10.99M to $11.19M. Selling less, earning more. Gross margin 35.7% to 50.0%.

[Stenly Tenev @StenlyTenev]

@wallstengine Home Depot's adjusted operating margin fell 10 bps. The reported one fell 19: 14.48% to 14.29%. $6.84B operating income on $47.86B of sales.

[Stenly Tenev @StenlyTenev]

Home Depot's August quarter operating margin: 14.3%.

Same quarter last year: 14.5%. Down 0.2 points.

The April quarter has been falling a full point a year, four years running. This one did not follow it. Anyone who read across from 11.9% and expected 10.9% was reading the wrong line.

I said I'd post the number either way.

$47.86B sales, $6.84B operating income. HD 8-K, quarter ended 2 August 2026.

[Stenly Tenev @StenlyTenev]

@danielnewmanUV Nvidia's April quarter ran a 65.6% operating margin, up from 49.1% a year earlier. The revenue line is the easy one to beat.

[Stenly Tenev @StenlyTenev]

@GrindeOptions Revenue per MW is only the numerator. SpaceX's June quarter still ran a -1.8% operating margin, on revenue up 91.9%.

[Stenly Tenev @StenlyTenev]

@zerohedge Off the balance sheet, not off the filings. Meta's $349.3B of contractual obligations is tagged in XBRL, and $81.7B of it comes due within a year.

[Stenly Tenev @StenlyTenev]

@amitisinvesting Items 1 and 11 are the same number. Ohio is 8GW total, phased from 2028. The model assumes 8GW every year.

[Stenly Tenev @StenlyTenev]

@Sam_Badawi June quarter operating margin, four years: 17.5%, 54.1%, 76.1%, 77.7%. The expansion added 36.6 points, then 22.0, then 1.6.

[Stenly Tenev @StenlyTenev]

@pdicarlotrader The March quarter swing was $140.4M of impairments. Revenue itself was flat to the dollar: $144.8M, against $144.8M a year earlier.