エンジニアリングフィールドノート。 Every backtest you have ever seen answers one question: what were the best settings for the past? A walk-forward conveyor answers the question that actually pays or costs money: if I re-tuned this strategy every week using only the data available at that moment, would the NEXT week have been profitable? We built that conveyor, fed it four widely published TradingView strategies, and ran 320 studies across 80 markets on two timeframes. This note is the full result — including the parts that lose money — and a practical map of what to do with it depending on who you are.

The four strategies

We deliberately chose strategies that are public, popular and reproducible — the kind a retail trader actually copies from a charting community:

  • PSAR flip — classic Parabolic SAR reversal: enter when the SAR dots flip sides.
  • Heikin-Ashi PSAR — the same SAR logic computed on smoothed Heikin-Ashi candles.
  • Dynamic RSI Momentum — an RSI whose length adapts to the correlation between price and its own momentum; buy the oversold cross, sell the overbought cross.
  • Pivot Reversal + trailing — breakouts of confirmed pivot levels, filtered by bar range in ATR units, with a trailing-stop exit.

Each entry rule, every parameter, and the exact search range and step for every knob are declared in code and published as a machine-readable passport. What is not written there is not searched — no quiet re-tuning after the fact.

How the conveyor judges a strategy

One cycle looks like this, and only ever uses data available before the week it trades:

  1. Train window (13 weeks on 4h, 2 weeks on 15m) — a genetic multi-objective search tunes entries and exits only inside this window.
  2. Stop at the Pareto front. Instead of crowning one "best" configuration, the search keeps the whole trade-off surface: highest profit, lowest drawdown, and every compromise between them.
  3. Hard constraints, fail-closed. Anything with a training drawdown over 30%, a negative training result, fewer than 8 trades, a weak profit factor, no stop-loss, or leverage above 3× is rejected before selection. If nothing survives, the week is traded 横ばい — an honest operator does not deploy the best of a bad lot.
  4. Five preset arms trade forward in parallel. From the surviving front we pick by five different rules — the balanced compromise, maximum profit, minimum drawdown, the "knee" of the front, and best profit factor — and each preset trades the next week independently. The forward week, not an opinion, decides which selection rule deserves trust.
  5. Slide one week, repeat. Old data ages out; the stitched forward curve is the only scoreboard, always compared against simply holding the coin.
training max drawdown → training net → min drawdown knee max profit
One training window's Pareto front. Green points survived the hard constraints; ringed points are presets that trade the next week in parallel. Grey points were rejected — too much drawdown, too few trades, or no stop.

What 320 studies actually said

On the 4-hour timeframe — the one with real statistical weight, roughly 200 sliding windows per market — the verdict is unanimous and brutal: all 60 studies, all four strategies, NO EDGE. Median alpha versus simply holding the coin ranged from −170 to −450 percentage points depending on the family. Weekly re-optimization, Pareto selection and disciplined risk limits did not save a single one. If you are running a Parabolic SAR bot on 4h crypto charts with any fixed or re-tuned settings, our data says the market has been charging you for the privilege.

On the 15-minute timeframe the picture is more interesting but younger: out of 260 studies, 42 finished above our noise floor — the best being a PSAR flip on VET with +39.5 points of alpha and both forward weeks positive. We do not celebrate these yet, and neither should you: with only 30 days of 15m history there are just two forward windows per study. The conveyor now re-runs automatically every week, and each week adds one more genuinely out-of-sample window. In four to six weeks these candidates will have either earned a real verdict or joined the graveyard. Both outcomes will be published.

The most transferable result is about method, not markets: among the five selection rules, picking the minimum-drawdown preset lost the least (median alpha −11.5 vs −13 for profit-seeking rules) — and no selection rule turned losing entries into winning ones. The edge, if it exists, lives in the entry logic. How you pick from the optimization front only decides how fast you bleed.

What should you do with this?

If you are new to trading

The single most valuable habit this research supports: never let a backtest screenshot make a decision for you. Start on paper, where every mechanism here is free to touch: create a demo account with a virtual balance at /app/start, watch honest strategies trade simulated money, and read the short lessons in the アカデミー — including why a 78% win rate can still lose money. When any platform (ours included) shows you a profit number, your first question should be: forward or fitted?

If you trade your own system

Everything in this pipeline is available to you as a research tool. Explore per-coin strategy evidence on the public research pages (for example /crypto/BTC), then sign in and open the optimizer to run honest in-sample/out-of-sample tests instead of single-split backtests. The twelve above-noise candidates from this campaign sit in the strategy rating grid as frozen, hash-locked presets marked paper-only — you can put any of them on a paper bot in two clicks and watch them earn or lose a track record in real time. Three rules we now consider non-negotiable: demand a stop-loss in every configuration, cap leverage at 3×, and judge every idea against buy-and-hold — not against zero.

If you follow signals

This is the audience with the most to lose. We once backtested a feed of 52,800 public crypto signals: following it faithfully would have destroyed the account. The four strategies in this study are better than that feed — real rules, reproducible logic — and they still failed the forward test on 4h. So before paying anyone for signals, ask for the one thing this article demonstrates: a stitched forward record with losing periods visible, not a highlight reel. Our own forward experiments — wins and failures alike — are published openly in the research lab and the running build changelog. If a seller cannot show you the equivalent, you are not buying a signal; you are buying a story.

If you are evaluating us as an investor or partner

Read this article as a due-diligence artifact. The system that produced it is designed so that we cannot quietly overclaim: verdicts are fail-closed (a strategy is guilty until forward data proves otherwise), paper track records are append-only and version-locked so losing trades cannot be edited out, and a campaign that tried 320 things has its significance bar raised 320-fold. Start with how we measure, then see 実際に販売されているもの — そして何が販売されているかに注意 ではない: この研究からは、収益を上げるボットとして販売されるものは何もありません。まだ何も稼いでいないからです。もしそのような規律をお探しなら、 お問い合わせください.

次に何が起こるか

コンベアは今や一度きりの調査ではなく、継続的なプロセスです。毎週、全15市場の4つの戦略すべてを新しいデータで再最適化し、各候補のフォワードレコードを1ウィンドウ分延長し、判定結果を再公開します。4時間足ファミリーは、実質的に新しいエントリー仮説がキューに入らない限り、エッジなしとしてクローズされます。候補が完全なラダーを越えると（十分なポジティブなフォワードウィンドウ、十分なトレード数、ノイズフロアを上回る）、それでもより長いペーパートライアルに進むだけです。ライブ自動化は、私たちにとってもプラットフォームを使用する誰にとっても、そのゲートの背後にロックされたままです。それはブランディングとしての慎重さではなく、この証拠に基づけば、単にデータが指示することです。