The research behind Muster

We tried to build a prediction engine. Here's what we found.

Before Muster, we spent weeks trying to build the thing most share apps imply they have: a model that reliably picks winners and times the entries and exits. We tested it the way a quant desk would – and we're publishing the results, failures included, because what we learned is the reason Muster only reports what you hold.

How we tested

372 ASX companies sampled across the whole market, deliberately spanning large caps down to micro caps – not just today's index leaders, because testing only the survivors flatters any strategy.
Ten years of daily data covering four separate market regimes: the 2016–19 grind, the 2020–21 pandemic, the 2022–23 rate shock, and the 2024–26 run.
Walk-forward evaluation – every decision uses only information available at the time. No peeking.
Realistic costs – brokerage plus the bid-ask spread on every trade, from about 0.26% a round trip on the big names to over 1% on the small ones. The single most common thing backtests forget.
Out-of-sample guards and significance thresholds, pre-registered before each run – a strategy had to work on data it was never tuned on, in every market regime separately.

The trap that killed most claims

Simply equal-weighting our sample of companies "beat" the index by nearly 18% a year. That looks spectacular and means nothing: it's the fingerprint of survivorship and small-company effects, not skill. So every strategy had to beat that same basket of companies, not the index. Held to that bar, the market-beating claims fell over one by one.

What we tested – and what the evidence said

NO EDGE AFTER COSTS

Trend and momentum entry/exit rules

Every weighting we tried was negative after costs out of sample. Tuning found rules that looked brilliant on past data and failed the moment they met new data.

FAILED IN EVERY CONFIG

Stop-loss and target tuning (nine configurations)

We swept the thresholds every reasonable way. Every single configuration lost money in the 2022–23 rate-shock window, and tighter stops were strictly worse. There was no setting that worked across all four regimes.

WORSE AS IT HARDENS

Market-regime timing (index above or below its 200-day average)

"Only hold when the market's trending up" sounds sensible. In practice it got monotonically worse the harder we applied it. It only ever did one useful thing – cut the pandemic drawdown roughly in half – and it charged about 5% a year of return to do it.

COLLAPSED VS THE FAIR BENCHMARK

Cross-sectional momentum and sector rotation

Ranking companies by recent strength, or rotating between sectors, looked promising against the index and evaporated against the same-universe basket – negative in at least one regime every time. Sector rotation lost about 4% a year outright.

PURE BID-ASK BOUNCE

Short-term mean reversion (buying the week's biggest fallers)

A promising lead vanished under scrutiny: the apparent gain was the bid-ask spread on illiquid losers, not a real, tradeable move. Delay the entry by a day or two and it turns sharply negative.

PRICED IN BY THE OPEN

Trading the ASX day on the overnight US move

The S&P 500's overnight direction is correlated with the ASX day – but the opening auction has already absorbed it. Enter at the open and the rest of the day drifts negative, even before costs. Decisively ruled out.

CALIBRATED & KEPT

Typical ranges

One thing the data genuinely supports: how much a company tends to move is forecastable even when the direction isn't. The "typical range this week" bands held the real outcome about 80% of the time across ten years of out-of-sample testing.

ACCURATE & KEPT

Describing what happened

The engine is good at what a rain gauge is good at: saying what is, clearly and consistently – what you hold, what it's worth, how it moved, what it paid. That is what Muster does.

Day trading is ruled out, structurally

This one isn't a judgement call. On the biggest, cheapest-to-trade ASX names, a day trade has to call the direction right roughly 68% of the time just to cover the round trip; on smaller names you'd need to capture the entire day's move. ASX returns accrue overnight, not during the session, and the one overnight lead that exists is gone by the opening auction. So Muster has no day-trading features, and won't.

What this means for Muster

We could have kept the buy/sell buttons and marketed the past-data numbers. Instead the findings are built into the product:

Reports, not predictions. Muster describes what you hold and how it moved. It doesn't forecast prices, and we don't pretend it can.
Honest ranges, not price targets. Calibrated "typical range" bands, because that's what the data actually supports.
Even the one survivor stays out. A single narrow momentum variant did survive our gates on the largest companies – as an unproven research candidate, not a promise. It isn't in the app, because "survived once" is not "works".
Nothing is traded, ever. Muster never touches your money and never connects to a broker.

Five questions to ask any stock-picking service

If an app or newsletter claims their picks make money, their evidence should survive these – ours is above, failures included:

Were trading costs included? (They erase most claimed edges on their own.)
Was it tested out of sample – on data the model was never tuned on?
How many companies, over how many market cycles? A handful over one good run proves nothing.
Does the universe include companies that were delisted, or only today's survivors?
Did it beat the same basket of companies, or just the index? Most things don't.

This page summarises internal research conducted with standard quantitative methods (walk-forward simulation, out-of-sample validation, significance testing). It is published for transparency and general information – it is not financial advice, and it makes no claims about any specific company. Share values go up and down; never risk money you cannot afford to lose.