One model that never sees the month it trades.
Pluto retrains on the last three months, trades the next one, then rolls forward — 43 consecutive months, 3,404 trades, every figure net of real spread, commission and slippage. A cross-asset engine watches the dollar, indices and crypto in the background. No cherry-picking, no look-ahead.
the ledger · every month, nothing hidden
29 green months out of 43. Including the red ones.
Each cell is one month, traded by a model that had never seen it. 2025–26 were never touched while tuning — the pure test: 13 of 19 green, +448% compounded. Hover a cell for detail.
| Year | Regime | Green | Avg / mo | Worst | Best | Year |
|---|---|---|---|---|---|---|
| 2023 | Gold bull leg | 10/12 | +10.1% | -2.5% | +52.3% | +185% |
| 2024 | Range, then breakout | 6/12 | +9.6% | -11.5% | +56.8% | +144% |
| 2025oos | Validation year — untouched during tuning | 8/12 | +10.9% | -15.0% | +49.3% | +199% |
| 2026oos | Validation + live deployment | 5/7 | +9.7% | -2.9% | +29.6% | +83% |
the universe · one bet, many eyes
It trades gold. It watches everything.
Tick history for 11 instruments is archived every week. The cross-asset engine already found a real, above-noise signal: gold’s next move leans on the dollar index and silver. It isn’t live yet — each passing month is a clean out-of-sample test before it ever touches a trade.
the traded market + its cousin
the dollar tells gold where it's going
risk-on vs risk-off flow
the macro and the fringe
method · why rolling beats remembering
Markets change their mind. So does Pluto — every month.
A fresh gradient-boosted model learns from the last 3 months of tick microstructure — 151 features per minute bar.
It trades exactly one month it has never seen. NY session, high-confidence signals only, hard risk rails.
The month is written to the ledger — win or lose. Nothing is retroactively fixed or re-run.
The finished month joins the training set, the oldest drops off, and the cycle repeats.
Built on a fixed engine, not a flattering one
Early versions of this backtest carried a subtle look-ahead bug: trades opened near a session close were priced against the next session. It inflated results by roughly 20×. We found it, fixed it, and rebuilt every number here with exits that walk the real, continuous price path.
The standard is simple: if a number can’t survive honest execution, it doesn’t get published.
Tested and rejected
Each of these was built, run across every month, and killed by the evidence. The survivor is what you see above.
- Static model (train once on 4 years)goes stale — break-even
- Expanding window (never forget)old regimes poison new months — worse
- 6-month rolling windowunstable — regime luck
- More indicators (+59 features)noise for 3-month windows — worse
- Fewer features (top-30 / top-60)loses the microstructure signal — worse
- 5-seed ensemblesbigger wins, not more green months
- Meta-labeling (AI judging the AI)AUC 0.50 — trade outcomes are unforecastable
- GRU neural sequence modeltrees beat nets on this data
- Direction prediction (any horizon)coin flip in 4 of 5 walk-forward years
- BTCUSD transplantcosts 3×, no session structure — rejected
risk · the rules that can’t be argued with
Seven hard rails. The model gets no vote.
a fresh model every month; last month joins the training set, the oldest drops off
13:00–16:00 UTC — where the edge was validated; flat at 15:59, never holds overnight
only high-conviction signals trade; whole months stand flat if quality is absent
high-confidence BUYs captured the bull months a short-only book missed
stops model-gone-wild days before they compound
one bad day can never become a catastrophic one
a hard ceiling — the moment equity falls 15% below its month peak, the bot stops for the rest of the month. Drawdown can never exceed it
position size cut to 40% at −1.5% from peak, 15% at −3% — losses shrink themselves before the cap is ever reached
read this before anything else
This is a high-variance system, published honestly. The worst single month was -15.0%; the deepest intra-month drawdown −15.1%. Past walk-forward performance — even honest walk-forward performance — does not guarantee future results. Pluto runs on demo capital. Nothing on this page is investment advice.
get access · own it or run it with us
Buy the system. We make it live for you.
One payment, no subscription. You don’t install or babysit anything — send us your MT5 login and password (demo or live), and we deploy Pluto on a dedicated server, retrain it every three months, and support it for life.
Want to see it run before you pay a rupee? Open a demo account with your broker, send us the demo id and password, and we’ll deploy Pluto live on our server and keep it running free for three months. No card, no commitment — watch it trade your demo and decide for yourself.
the exact rolling 3-month model on this page, with all seven risk rails built in
send us your MT5 login + password (demo or live) and we run it for you on a dedicated server — nothing to install
we keep the model fresh on a rolling window, on schedule, for as long as you run it
one-time payment, no subscription; questions and maintenance covered for life
reach out and tell us your broker and account type
your MT5 id + password (demo first is recommended)
the bot goes live on our server, wired to your account
retrained every 3 months, supported for life
lifetime access
contact
bhargavsuhagiya99@gmail.comTo join, invest, or ask anything — just send a mail. Demo capital strongly recommended before live. High-variance system · not investment advice.