One model. Every trade recorded since January 2008. Your opportunity to follow the research desk.

Model Portfolio, whole shares, anchored 2 January 2008 at $100,000 USD, as of the 4 September 2026 close. All figures in US dollars. Based on backtested results. Past performance, whether actual or backtested, does not indicate future results.

Research from a model that shows the receipts

Every decision, every day, verified before it reaches you.

Most investment research asks you to trust the analyst. Kodama's research is different in one important way: the analyst is a rules-based quant model, and every trade it makes is shown in full. Each trading day the Kodama Research Desk publishes what the model did, what it is holding now, and how the record is adding up, in the same way a good analyst publishes ratings and stands behind them. The difference is that this analyst has no opinions, has locked rules in place to dictate trades, and has every decision since 2008 on the record for anyone to check.

That is what a subscription is: a seat at the desk while a disciplined, tested model works, with every trade on the table and nothing dressed up. The desk monitors the model independently and verifies every trade against the ledger before it is published. What the desk never does is change the rules, override a decision or add a view of its own. The rules were locked before the record began, and the research is the honest, daily publication of what they produce. How that research fits your own thinking is, as with any research, yours to decide.

The simplest way to picture it is a cook working through a dish on television. You watch every step as it happens: what goes in, when, how much of it, and what comes out at the end. The desk works the same way. You see each position as it opens, the weight it was given, when it was trimmed, when it was closed, and how the desk handles the ordinary business of running a book. What you do not get is the recipe. The measurements behind the rules stay with us. That is the only part held back, and it is the part a subscription is actually buying.

Most systems talk about potential. This one has been through five declines.

A strategy that has only ever run in a rising market has not been tested, it has been lucky. This one has been run through every major decline since the anchor: 2008, 2018, 2020, 2022 and 2025. Not as a hypothetical, and not on rules invented afterwards. The same rules, on the same pinned data, through each of them.

It is not a magic trick. In a sharp panic the model can fall as hard as the market or harder, because it holds individual companies rather than five hundred of them. What it does is recover faster and take more of the upside when the recovery comes, and that combination is what compounds. The research desk measured it directly, because a strategy that only looks good on the way up is not a strategy.

Capture ratios, positive months and the best and worst months are computed on 224 monthly observations since January 2008. Volatility and Sharpe annualised from daily returns, risk-free rate zero. As of the 4 September 2026 close. Based on backtested results.

Taken together, those boxes are one argument. The model is more volatile than the index, by about a fifth, and that is the honest cost of concentration. The question the desk cared about was whether the return justified it. Roughly one and a half times the market's gains, roughly three quarters of its losses, an annualised 30.96% against the index's 11.31%, and a Sharpe of 1.25 against the index's 0.64 says it did. The return pays for the extra movement rather than merely absorbing it.

What it actually is

In one paragraph, before any jargon.

Bonsai Nobi buys individual stocks as they recover an uptrend, sizes each position by the strength of that particular setup, and sells on a rule rather than on an opinion. When it is not holding stocks it holds the market itself. It trades a curated pool of large American companies, it is long only, it uses no leverage and no options, and it makes one decision a day, after the close. It does not predict anything. It reacts to what price has already done.

How a trade happens

Four steps. The measurements behind them are the part we do not publish.

Most of the work was throwing things away

Hundreds of strategies were built and tested before this one was kept. Almost all of them died for the same two reasons. Either they were not consistent, producing a good year and then a terrible one on the same rules, or they only worked in one kind of market, which is the same failure wearing a better disguise. A rule that prints money in a bull run and hands it all back in a bear market has not found an edge, it has found a trend.

The survivors were the ideas that still worked when the conditions changed, and that kept working when combined with each other. The discarded tests were kept as well, with the reasoning written down before the results were seen, which is why the archive can be shown to anyone who asks how a rule earned its place.

It hunts. And it hunts in a curated wood.

The rules matter. Where you point them matters just as much.

The same rules do not behave the same way on every stock. Some companies move in the clean, repeatable way this system is built to read. Others are noisy, gap-prone or news-driven, and they generate entries that look right and end badly. So the universe is not the whole market. The Kodama Research Desk starts from the S&P 1500 and curates from it, keeping only the names where the strategy has demonstrated a genuine risk and return relationship rather than a lucky run.

Two measures drive that selection more than raw return. Win rate, because a strategy you cannot stay with is worthless no matter what the curve says, and profit factor, because what actually compounds is the relationship between what the winners make and what the losers cost. A name that returns well but only through two enormous outliers is a name this system does not want.

Closed single-name positions in the Model Portfolio record, anchored 2 January 2008, as of the 4 September 2026 close. One position counts once: partial sales taken along the way are aggregated into the position they belong to, not counted as separate trades. Returns are measured against the position's original cost. Based on backtested results.

Read the third box before the first. Being right about two thirds of the time is good but not extraordinary. What matters is that the average winner is close to twice the size of the average loser. That asymmetry is the engine, and it is what the exit rule exists to protect: losers are cut while they are still small, winners are left alone to become large.

One note on how these are counted, because it is the easiest place in this industry to flatter yourself. The model sometimes sells part of a position while keeping the rest, and by rule it only ever does that to a winner. Counting each of those partial sales as its own trade would lift the win rate to 75% and it would not mean anything, because you would be counting the same good position several times. So a position counts once here, from entry to final exit, with everything realised along the way folded into it. That is the conservative reading and it is the one we publish.

The list is not set and forgotten

Reliability is not a permanent property. A company changes, its sector changes, the way it trades changes, and a name that suited this system for six years can stop suiting it. So the desk reviews the universe continuously and acts on what the review finds. Names are promoted when they earn it, demoted when their reliability slips, and removed when they no longer belong. New candidates are watched and measured before a single dollar goes near them.

Reviews happen on a schedule and recommendations are logged, but changes take effect only at set reselection points rather than the moment someone has an opinion. That gap is deliberate. It is the difference between a curated list and a reactive one, and it is the discipline that stops a bad month from rewriting the roster.

A complete record, not a tip

Most services publish a name and stop. That is the easy quarter of the picture, and on its own it is close to useless. A name without the size the model gave it, the moment it acted, and the rule that later took it off the book is not a record of anything. So the research reports all four, for every position, every time.

All four, every time, published together. What you are reading is a record of decisions already made by a set of rules, not a suggestion about what you should do next.

The record, year by year

Every calendar year since the anchor, against SPY. Nothing removed.

2026 is year to date at 4 September 2026. Model returns are the whole-share Model Portfolio track. Based on backtested results.

Two losing years in nineteen, and in both of them the market lost more. The model does not avoid bad years. It finishes them in better shape, and it does not need a heroic recovery to get back to where it was.

What it will and will not do

The second column matters more than the first.

The question the desk kept asking: what if you buy at the worst possible moment?

Rolling three-year and five-year windows, every single starting day.

A single since-inception number tells you almost nothing on its own, because nobody reads a record from its first day. People start paying attention on some ordinary Tuesday, possibly the worst one. So the research desk did not judge this on the headline. It judged it on every three-year and five-year window in the record, one for each trading day, which is 3,945 three-year windows and 3,444 five-year windows.

That is a deliberately unkind test. It includes the windows that begin the day before a crash, and the ones that begin at a market top. It is also the only test that answers the question a careful reader actually has, which is not "how did it do since 2008" but "how did it do from a bad starting point".

Every rolling window with a complete history in the record, measured daily, anchored 2 January 2008, as of the 4 September 2026 close. The large figure is the total return over the window; the annualised figure beside it is the constant yearly rate that compounds to it. Based on backtested results. Past performance, whether actual or backtested, does not indicate future results.

The worst number in that table is the one worth sitting with. There is no three-year window in the record where the model lost money, and the weakest one still compounded at roughly 6% a year. The market has three-year windows that finish underwater. Across five years the gap widens rather than narrows, which is what you would expect from an edge rather than from luck: luck decays with time, and a rule set does not.

Built by a trader. Checked by a professional.

The rules came out of one trader's decade of work, and they were shaped in consultation with a professional portfolio manager who runs real client money for a living. That mattered in a specific way. A private trader optimises for return. Someone answerable to clients has to care about the things that make a strategy survivable: how deep the losses go, how long they last, whether the thing can be followed by a person with a job, and whether the record can be defended line by line when someone asks.

So the design questions were not only "does this make money". They were "would a client stay invested through this", and "can every line of this be defended when somebody asks". That is why the exit rule exists, why the position sizes are graded, and why the losing tests were kept.

Why the rules are not published

You will notice this page names the triggers but not their settings. That is deliberate and it is the honest trade. The entry threshold, the exit rule's parameters, the grading formula and the funding logic are the decade of work, and they are the only thing a subscriber is actually paying for. Everything downstream of them, the trades, the timing, the sizing and the results, is published in full, every trading day, along with how the desk handles the ordinary running of a book. The steps are on show. The measurements are not.

What you can verify instead is the record. It runs on a single continuously invested portfolio anchored in January 2008 that has never been restarted, re-seeded or re-anchored, on pinned data that is appended to and never rewritten, with every rule change registered in writing before its results were seen. The archive includes the tests that failed. That is the part a competitor cannot copy, because nobody fabricates their own losses.

Questions