Ozmium DocsSignal Methodology
How Ozwald reaches a trading call - the inputs he reads, the parts that stay proprietary, and the evidence standard every published signal has to clear first.
Ozwald uses this method to size his own positions and to decide what the news desk covers. The published signals are a by-product of that, not a separate product.
What Goes In
Four inputs feed every decision. They frequently disagree.
| Input | What It Contributes |
|---|---|
| Market structure | Price, range, volatility and momentum from one minute to one day. Each timeframe is judged on its own closed bars; a forming candle is never read |
| Onchain data | Positioning from the venue: which side the open interest sits on, order flow, and the size actually at risk |
| The news desk | The same pipeline that writes the bulletins, so what is happening to an asset feeds into whether its chart means anything |
| The committee | A proprietary ensemble that weighs the other three and decides whether their disagreement is fatal |
What Stays Private
The specific thresholds, the weighting between inputs, the per-timeframe gates and the composition of the committee are not published.
The strategy is proprietary. The standard of proof it has to clear is not, and it is described in full below.
What Every Signal Has To Clear
Grading is causal. Entry is the next bar's open, not the close that produced the signal. The stop is checked before the target, so a bar that touches both is recorded as a loss. Fees come off every result at the venue's published rate for that pair, not an average.
Results are re-measured out of sample. Every candidate is tested on four disjoint splits: two halves of the market list, and two halves of the time window. Failing any one of the four discards the candidate.
Each asset class is measured on its own. Gold, semiconductors and tokens do not share a volatility profile, so a bracket fitted on one is not applied to the others.
Failing timeframes are not switched off. They keep generating calls into a shadow ledger, graded the same way and shown to nobody. A frame promotes itself when its own live record clears the bar and demotes itself when the record decays. Neither happens on a small number of trades.
Three Times This Caught Us
Three errors from one week of development, all found by re-measuring rather than by review.
A twelve-market panel showed a profit factor of 1.22 on a fast timeframe, and the curve was monotonic - the number rose every time the filter tightened. Re-running the same test on twenty-five markets gave 0.97. The curve had been correlation between markets rather than edge.
One shared bracket across all asset classes let whichever class had the most data set the stop distance for the rest, which made a profitable timeframe look unshippable. The question was wrong before the number was.
Weekend gap handling introduced its own error. Filling stops at the next open is correct for a market that closes and wrong for one that never does. It charged round-the-clock markets for gaps they never had and pushed a paying frame below break even. Gap handling now keys off an actual break in the tape.
Checking Our Work
GET /api/signal-proof returns the derivation behind any call: the bars it read, a fingerprint of
that data, each indicator's value at that moment, and the textbook formula for each one. The
individual reads are standard technical analysis.
The endpoint does not return how those reads are combined. Every input is verifiable and every number reproducible; the ensemble is not published.
What We Do Not Claim
Signals are commentary and infotainment. They are not financial advice and not a promise. A measured profit factor describes a past sample and does not forecast your account.
Ozwald trades his own book on this method.