We tried to beat gold.
We failed honestly.

GoldWall Analytics is a solo-run quantitative research lab focused on a single instrument: gold against the US dollar (XAUUSD). We are not a broker, a fund, a signal service, or a licensed adviser. We publish measurements — a live volatility-regime reading, institutional positioning from the CFTC Commitment of Traders report, and the macro backdrop of real yields and the dollar.

The work began as an attempt to build a profitable automated trading system. After a long sequence of rigorously tested strategies all failed the same honesty bar, we did the thing most research shops quietly avoid: we accepted the negative result and pivoted from trading the market to measuring it. This site is the measurement layer that survived.

A backtest that looks profitable is not evidence. It is a hypothesis that has not yet been attacked hard enough. Before any number earns a place on this site, it has to survive four tests at once:

• Real costs. Spread and commission are charged on every simulated trade. An edge that only exists gross of costs does not exist.
• Chronological, out-of-sample splits. Models are trained on the past and tested on a future they never saw. No shuffling, no peeking, no look-ahead.
• Honest labels. One shared feature-and-labeling implementation is used across training, verification, backtesting and live serving — never a second copy that quietly drifts. This single rule caught train/serve skew again and again.
• A concentration & regime check. We ask the question that kills most strategies: does the profit come from the whole period, or from one lucky week or one unusual market regime? If the equity curve is one spike wearing a trend-coat, it dies.

A positive result triggers more skepticism, not celebration. And we never tune a strategy until it becomes profitable — that is just overfitting a signal that was never there.

Almost everything here rests on walk-forward validation. Instead of scoring a model on the same history it learned from, we march through time in folds: train on a window, test on the next unseen window, step forward, and repeat. The model is judged only on data that did not exist when it was fit.

We then look at the folds individually, not just the average. A strategy that wins on average but only because of a single dominant fold has not found a durable edge — it has found one market it happened to fit. Averages flatter; the fold-by-fold view is where honesty lives.

Between the theoretical edge in a price series and a solo retail trader stands a wall of spread, commission and market efficiency. We ran into it eight times. We are keeping the scorecard public because the failures are more valuable — and more honest — than any of the survivors:

• Order-book microstructure (direction). An early ensemble looked predictive; the old code had fatal bugs and the signal was illusory once fixed.
• Richer order-flow features (MLP & gradient-boosted). A genuine ~54% directional ceiling at a 60-second horizon — entirely eaten by transaction costs.
• Micro-scalping. Structurally rigged against retail by spread plus commission before any skill is applied.
• Indicator-based direction on 5-minute bars. Flat, coin-flip probabilities out of sample. No signal at all.
• Volatility-breakout trading. Up ~$20k over twelve cherry-picked months; down ~$7.7k across the full multi-year sample. That first number was regime luck.
• EUR/USD mean-reversion. A statistical break-even — the signature of an efficient market with nothing left on the table for us.
• H1 swing trend-following. A tempting +$8.9k — until we found 86% of it came from a single week. Killed by the concentration check.
• The same swing model with fixed-risk stops. Down roughly 90% of the account: fixed stops became the width of ordinary noise.

Adding exogenous data (the dollar index, real yields) generally made things worse. Only the Commitment of Traders series carried genuine information — and even that meaningfully helped in just one bearish stretch. Taken together, these results are a careful, expensive-to-obtain measurement of just how efficient the gold market is for a solo participant. We consider that a finding, not an embarrassment.

One model cleared the bar. Not a direction predictor — a volatility-expansion classifier. It estimates the probability that gold makes a large move over a defined forward horizon, and it beat a naive baseline in every out-of-sample fold and across bull, bear and range regimes. It says nothing about which way price will go.

That distinction is the whole point. Predictable volatility is not the same as profit. Knowing a storm is likely is useful context; it is not a trade. So we publish the storm forecast and refuse to dress it up as a direction call.

Read this section as the load-bearing one. Everything on this site falls into two honest categories:

• The volatility-regime reading is a mathematical probability from a validated statistical model — an estimate of the chance of a large move, with a known method and known limits. It is not a prediction of direction and not a guarantee of anything.
• The COT and positioning figures are a sentiment read: a transparent transformation of public CFTC data into a 52-week context (the COT Index). They describe how crowded a side of the market is. They are not a timing signal.

Neither is financial advice, a recommendation, or a solicitation to trade any instrument. Nothing here accounts for your circumstances, and no output should be treated as a prompt to buy or sell. Markets carry substantial risk; past behavior does not guarantee future results. See our Terms for the full disclaimer and our Weekly COT Report for the positioning data in long form.

We are explicit about where every number comes from and how fresh it is:

• Price / volatility — hourly XAUUSD candles via the cTrader Open API, spanning multiple years; the regime model refreshes on a short cadence off the request path.
• Positioning — the CFTC Commitment of Traders report, published weekly, with our 52-week COT Index computed on top.
• Macro — real yields and related series from FRED, and the US dollar index from ICE, aligned daily with rolling correlations.

Caches are refreshed on a schedule rather than on your visit, so the page stays fast while the data stays current. When a source is temporarily unavailable, the affected panel says so plainly rather than showing a stale or invented number.

Radical honesty includes being correctable. If you believe a number here is wrong, or a method is flawed, tell us and we will investigate and publish a correction where warranted. Reach the lab at contact@goldwallanalytics.com.