Choose an institutional-grade systematic strategy — or let six questions choose it for you — then press go. Your own Blackbox starts trading on paper, marked to live markets, guided by validated rules and an AI that reads the news.
Answer six questions and every answer visibly shapes the outcome — or skip the quiz and build your own mix straight from the numbers.
A six-question profile. We show exactly which answers capped or shaped your mix — no black box.
Browse the strategies below and set your own weights by CAGR, volatility and max drawdown.
Every stat is a 20-year backtest — true drawdowns included. The AI scanner's live read is on every card. Click a card for the full profile.
Set weights with the steppers on the strategy cards below. Your mix must total 100%.
This page is the heart of the project. Blackbox runs four distinct AI systems — but the one that matters most is Regime Radar: five machine-learning models we designed and trained ourselves, from raw market data. The others read the live world and explain it. Below you'll find every system, each with a plain-English "how it works" — and further down, an in-depth account of exactly how the models were built, what data they use, and how they learn.
Five neural networks we trained to forecast market turbulence.
our own modelAI reads live news & social mood across six asset classes.
reads the worldAI ties today's news to the closest real past market event.
reads the worldAI explains how it's supporting each strategy right now.
explainsThis is what we built and trained ourselves — five neural networks, one per asset class, not a borrowed language model. Each learned from ~20 years of real daily market data to forecast whether the days ahead will be turbulent (high volatility) or calm. Every other AI feature in Blackbox uses a pre-trained model; these we trained. Here are their live calls and honest, measured track records — including which market's turbulence turns out to be most predictable.
This is the long version — everything you'd want to know about the machine-learning core. What the models do, exactly which data they learn from, how they were created, and how a neural network actually teaches itself. No hand-waving.
Financial markets swing between calm stretches and turbulent ones — periods of large, rapid price moves. Turbulence is when investors get hurt: crashes, whipsaws, forced selling. So we built models to answer one precise question, separately for each kind of market: will the next ten trading days be turbulent (unusually high volatility) or calm? — using only information available today.
We deliberately forecast volatility rather than direction (whether prices go up or down). Direction is very close to impossible to predict — markets are highly efficient, so a direction model is essentially a coin-flip and would make a dishonest demo. Volatility is different: it clusters. Turbulent days tend to be followed by more turbulent days, and calm by calm. That clustering is a real, well-documented statistical property, which means volatility is genuinely — if imperfectly — learnable. Choosing a target a model can actually learn is itself part of doing honest machine learning.
The models never buy or sell anything. They output a risk signal: a probability of turbulence ahead, which becomes a simple Calm / Turbulent call. In the wider product this could tell a portfolio to carry less risk before a likely storm — but the trading decisions themselves always stay with the validated, rule-based strategies.
A model is only as good as what it learns from, and the challenge specifically rewards working with a large dataset. We assembled five, one per asset class, from real daily market history — over 220,000 labelled examples in total. Each row is one real trading day for one real instrument, described by the features in the next section and labelled by what actually happened next.
| Model | What it learns from | History | Examples |
|---|---|---|---|
| Cataclysm crypto | 13 coins — Bitcoin, Ethereum, Solana, Cardano, XRP, Dogecoin, Chainlink, Litecoin, Bitcoin Cash, Ethereum Classic, Avalanche, Polkadot, Polygon | 2016–2026 Coinbase (free) | 29,600 |
| Vantage equities | 8 global stock indices — S&P 500, Nasdaq 100, Dow 30, DAX 40, FTSE 100, Nikkei 225, Euro Stoxx 50, ASX 200 | 2006–2026 OANDA | 40,700 |
| Bastion bonds & rates | 6 government-bond instruments — US 2/5/10/30-year, German 10-year (Bund), UK 10-year (Gilt) | 2006–2026 OANDA | 30,400 |
| Crucible commodities | 12 markets — gold, silver, platinum, palladium, copper, WTI oil, Brent oil, natural gas, corn, wheat, soybeans, sugar | 2005–2026 OANDA | 63,300 |
| Meridian FX | 10 currency pairs — EUR/USD, GBP/USD, USD/JPY, AUD/USD, USD/CAD, USD/CHF, NZD/USD, EUR/GBP, EUR/JPY, GBP/JPY | 2005–2026 OANDA | 57,000 |
Crypto history comes from Coinbase's public API (free, open); the four traditional-asset classes use OANDA, a regulated broker whose data feed gives clean daily prices going back twenty years. We pool every instrument within a class — each coin, index or pair contributes its whole history as training examples — which is what multiplies a few thousand days into tens of thousands of examples and gives each model enough to learn from.
A model cannot read a price chart the way a person does. It reads numbers. So for every single day, we compute 15 numbers ("features") that describe how that market had been behaving up to that point. These are the model's senses. They fall into five families:
One rule governs all of them and it is the most important rule in the whole project: every feature uses only information from the past. The feature for a given day is computed strictly from prices up to and including that day — never after. If even a hint of the future leaked into the inputs, the model would look brilliant on paper and be worthless in reality. Getting this "no lookahead" discipline right is what separates honest financial machine learning from the kind that quietly cheats.
To teach a model, every example needs a correct answer to learn from. For each day we look at the future — the realised volatility over the following ten trading days — and mark the day turbulent if that future volatility lands in the top third of what that market experienced during the training era, and calm otherwise. Crucially, the cut-off that defines "top third" is measured only on the training data, then applied unchanged to later data, so the model is never tuned to the period it is tested on.
The model at the heart of each class is a neural network — the exact kind of "self-learning algorithm" the challenge names. Here is what that means, concretely.
A neural network is a web of simple maths units called neurons, arranged in layers. Ours has an input layer of the 15 features, two hidden layers (of 32 and then 16 neurons), and a single output that becomes the probability of turbulence. Every connection between neurons carries a weight — a number that says how much that input matters. A network with these layers has a few thousand of these weights.
At the start, all those weights are random — the network knows nothing and guesses blindly. Then the learning happens, and this is the part that makes it "AI": we show it a training example (15 features → the true calm/turbulent answer), it makes a guess, we measure how wrong the guess was, and the network nudges every weight a tiny amount in the direction that would have made the guess less wrong. This nudging procedure is called backpropagation with gradient descent. We repeat it across all the training examples, over and over (each full pass is an "epoch"), and with every pass the network gets a little better at mapping "what the market looks like today" to "turbulent or calm ahead". After enough passes it stops improving, and we stop training automatically (this is "early stopping", which prevents it from simply memorising). At no point does a human write the rule — the network discovers the relationship in the data by itself.
A few practical touches make it work honestly and well: we standardise the features (rescale them to a common range) so no single one dominates just because its numbers are bigger; we balance the classes so the model doesn't lazily learn to always say "calm" just because calm days are more common; and we tune the exact probability cut-off (how sure it must be before it says "turbulent") on a separate validation set, never on the final test. As a cross-check we also train a random forest — a completely different model made of hundreds of decision trees that vote — and keep whichever of the two does better on validation. For four of the five classes the neural network wins or ties, so that is what we deploy.
The single easiest way to fool yourself in machine learning is to test a model on data it already studied. So we split each dataset by time, never randomly: the model trains on the oldest ~60% of history, we tune choices on the next ~15% (validation), and we grade it once on the most recent ~25% — data it never saw during training. Splitting by time matters because shuffling the days would let tomorrow leak into yesterday.
We grade with a metric called AUC, which runs from 0.50 (a coin-flip — no skill at all) to 1.00 (perfect). AUC is the right choice here because it isn't fooled by the fact that calm days outnumber turbulent ones. We always compare against two deliberately dumb baselines: "always guess calm", and the strong one — "volatility is already high, so expect more" (persistence). Beating persistence is the real test, because volatility clustering makes it a genuinely tough benchmark.
On data none of the models had ever seen, every one beats both baselines. Bond and rate turbulence turned out the most forecastable (AUC 0.93), then commodities (0.85), equities (0.82) and FX (0.81); crypto — the wildest, most headline-driven market — is the hardest (0.70). The live scorecard above shows all of this, including the edge column, which is the honest measure of what our machine learning adds beyond simply noticing that volatility is already high.
Running one identical method across five markets produced a genuine cross-asset insight: the steadier the market, the more predictable its turbulence. That makes economic sense — bond volatility is driven by slow, telegraphed forces like the interest-rate cycle, whereas crypto turbulence is often set off by sudden, unforecastable news. Discovering that by experiment, rather than assuming it, is exactly what a good study should do.
Once trained, each model is saved to disk as a small file — its learned "brain". When you open this page, the app takes the most recent real prices for a representative market in each class (Bitcoin, the S&P 500, the US 10-year, gold, EUR/USD), computes today's 15 features, feeds them through the trained network, and reads out a fresh turbulence probability. That is the Calm / Turbulent call you see at the top of the page, updated from live data.
We say plainly what these models are not. They are a modest, real edge — not a crystal ball. They miss sudden shocks that have no footprint in past prices, they raise occasional false alarms, and the high headline AUCs owe a lot to volatility simply being persistent (the model's own contribution is the smaller "edge"). They forecast how bumpy the ride will be, not which way prices go, and they inform risk rather than making a single trade. Every one of those caveats is stated because a model you can honestly critique is worth far more than an impressive number you cannot defend — which, in the end, is the whole point of the project.
An AI engine scans hundreds of news and social sources, scores each asset class, explains itself — and trims exposure in your Blackbox when it turns strongly bearish on a strategy's assets.
AI reads today's real headlines and matches them to the closest event from a fixed library of real past market events — then shows how each strategy's real, backtested returns actually moved in the months after. AI only picks which past event fits; every number comes from the same 20-year backtests as the rest of the app.
Each strategy's position is decided by its validated rule. On top of that, the AI reads the live news sentiment for that strategy's assets and acts as a risk brake — trimming exposure to 75% when the mood turns strongly bearish. Below, AI explains each strategy's current stance and how the overlay is supporting it. The rule is the driver; the AI is the co-pilot.
Every cycle, the AI scanner reads hundreds of real news and social sources and scores six asset classes. This briefing synthesises that live read — plus the closest historical analog — into a plain-English summary of the day. It describes the mood and what it means for a diversified systematic portfolio; it is not advice, and it never invents a number.
Systematic investing has its own vocabulary. Tap any term below for a plain-English explanation from the AI, or ask the tutor anything — how a strategy works, what a metric means, why diversification lowers risk. It's grounded in this app's own real strategies, and it never gives buy/sell advice.
Blackbox is built on eight strategies validated over 20 years of real market data. Artificial intelligence is layered through the whole product — but in a very deliberate way. This page explains exactly what the AI does, how it's implemented, and, just as importantly, what it does not do.
Every buy, sell and hold decision comes from a validated, backtested rule — not from AI. The AI reads the market, supports those rules with a bounded risk brake, personalises your mix within strict limits, and explains everything in plain language. It never trades on its own, and it never invents a number.
There is one crucial distinction on this page. Every feature below uses a large language model that someone else pre-trained. Regime Radar is different — these are models we built and trained ourselves. We took ~20 years of real daily market data across five asset classes (crypto, equities, bonds, commodities, FX — over 200,000 examples in total), engineered the input features ourselves, defined the target (will the next 10 days be turbulent or calm?), and trained a separate neural network for each market to learn the pattern — no human wrote the rule, the models discovered it from the data. We then tested them honestly on years of data they never saw during training: they score AUC from 0.70 (crypto) up to 0.93 (bonds) — genuine, if imperfect, skill that beats the naive baselines everywhere (a coin-flip is 0.50). This is the machine-learning core of the project — see all five live on the Intelligence page, with the honest results, cross-asset comparison and limitations in our written reflection.
Every cycle, an AI model reads hundreds of real news headlines and social posts across six asset classes — equities, gold, rates, FX, crypto, oil — and scores the mood from −100 to +100, with a plain-English explanation traceable to the actual headlines. It reads for real meaning (handling context and negation), not just keywords.
AI matches today's real headlines to the closest event from a fixed library of real past market events — the 2008 crash, the Taper Tantrum, COVID, and more — then shows how each strategy's real, backtested returns actually moved in the months after. AI only picks which past event fits; every number is from the backtests.
This is where AI directly supports each strategy. When the scanner reads a strategy's assets as strongly bearish (a score below −40), that strategy's exposure is automatically trimmed to 75% — a risk brake layered on top of the validated rule. The rule decides the position; the AI decides whether to ease off it.
For every strategy, the AI reads its live rule-based stance and the live sentiment on its assets, then writes a short note explaining where it's positioned right now and whether the risk overlay is supporting or trimming it. You can see this on the Intelligence page.
After the six questions, you can describe your situation in your own words. AI reads it and may apply a small, bounded adjustment to your mix — never more than a gentle tilt, never adding or removing a strategy, never overriding the validated math. And it always explains, in writing, what in your words drove the change.
Once your portfolio is live, AI reviews your actual mix — how diversified or concentrated it is, and its blend of return and risk — reading the exact figures for you. The blended numbers are computed precisely in code; the AI only puts them into words.
A plain-English summary of the day's market, synthesised from the live sentiment scan and the historical analog.
Your Blackbox writes its own summary, and you can ask it questions — answered only from your real portfolio data.
An AI tutor answers investing questions grounded in this app's own strategies, and any metric can be explained on tap.
AI explains why any strategy works, and talks you through the trade-offs when you compare two — always from the real numbers.
All of the above runs through a single shared connector to a large language model (Llama, served by Groq), behind one API key. Every AI feature "fails open": if the model is ever unavailable, the app keeps working and the feature simply says so — the validated strategies and every real number are never affected. That's the whole design: real, tested rules at the core; AI as the analyst, risk co-pilot, and teacher around them.
Pick any two of the eight strategies to compare their real 20-year statistics head to head — return, risk, drawdown and how independent each is from the stock market — then let the AI talk you through the trade-offs. Every number is from the same backtests as the rest of the app.
Log back in from any device. Practice capital only — no real money.
This key is your login. Save it now — it's shown only once.