Strategy research engine
A self-correcting pipeline: the constructor generates edges → walk-forward selects them → the forward gate confirms them live → only the proven ones reach trading with risk-based sizing
Laboratory + walk-forward
Thousands of strategy combinations are tested with a 70/30 in-sample / out-of-sample split and a bootstrap confidence interval on the profit factor — protection against overfitting.
Every horizon: M5 → W
Minutes, hours, days and weeks in parallel. A continuous flow of trades: the fast ones capture frequency, the slow ones hold the trend.
Barbell: runner plus small wins
Convex runner trades (a trailing stop with unlimited upside, catching the huge trends) alongside frequent small gains. A balance of frequency and force.
Level-based analysis
Support and resistance, level rejection and breakout-retest, break of structure, and multi-timeframe entry along the higher-timeframe trend — rather than a sum of indicators.
Honest costs
The backtest accounts for spread, slippage and overnight financing. A strategy passes only if it is profitable after every real broker cost.
Laddered take-profit
A mandatory target on every trade: it starts small and grows as wins accumulate. Profit is banked rather than handed back.
Reassessment and profit protection
After each winning trade comes a pause and a fresh read of the market, so nothing is entered on stale momentum. If a profitable session starts giving back, trading stops until tomorrow.
Promotion to live
Strategies first prove their edge in forward testing (£200 virtual, real prices). Only the winners are released to real money; those that degrade are withdrawn.
Local AI and the AI gate
A neural network trained on trading literature runs in shadow mode and earns its way to live: only on a proven edge after costs does the system release it into real trading.
Parallel search
The lab runs in 10 parallel workers and refreshes the strategy set daily — the engine keeps hunting for new edges on fresh data.
Account protection
A daily stop-loss, a minimum position size, a correlation limit and margin control — protection against forced closeouts and losing streaks.
Multi-broker
OANDA, IG, Saxo, Binance, Trading 212 — one engine executes strategies across several accounts with a SIM ⇄ LIVE switch.
Tiered capital system
As the account grows the engine raises the number of simultaneous strategies and positions (in £100 tiers) while keeping trade size fixed — more diversification without more risk per trade.
Profit protection ratchet
The day's profit is locked in steps: give back 10% from the peak and that level becomes a floor which is never surrendered; the bot then re-reads the market and carries on. Profit only moves upward.
Weekend protection
Before Friday's close, profitable positions are booked while losing ones are not closed but given a tight stop (additional loss ≤ £0.10) — protection against weekend gaps and swap. Round-the-clock crypto is left alone.
Breakout on volume plus sweep
Entry on a confirmed level break with elevated volume (accumulation, then impulse) with liquidity taken before the move — not on a random signal. A trend filter prevents catching a falling knife.
Movement to a real target
Trades are managed by server-side target, stop and trailing against ATR-based objectives rather than cut at spread scale — spread and commission stop eating the profit on short moves.
Proven edges from research
We take only documented edges (time-series momentum, Connors' RSI(2), Gerchik momentum) and test them in our own walk-forward lab on our instruments after costs. No holy grails taken on trust.
Profit runs to target
A winner is carried to its reward-to-risk target rather than clipped for pennies, exactly as manual trading from a level would do. Capital protection sits at the level of the day (a peak ratchet plus a daily stop) instead of strangling individual running trades.
Diversification across edges
Each edge on the market that suits it: trend on gold, mean reversion on indices, levels on the yen and metals. When one market stalls another works, and the portfolio holds up better.
Observability and alerts
Silent failures are caught in every module, a health monitor runs every 5 minutes with an email alert on breakage, logs rotate on a two-week cycle, and daily snapshots allow recovery from a major failure.
Edge constructor — 516 recipes
A strategy factory: from a library of primitives (43 triggers × 12 regime filters) it generates 516 candidate recipes, including candlestick figures (engulfing, pin bar, star, divergence, narrow-range break). The lab runs them through walk-forward and the null gate across the whole instrument universe and keeps only the survivors — the engine invents new edges every night.
Forward gate: the second layer
Between “the lab found it” and “we trade it at full size” sits a live forward test — data that arrived after the discovery and cannot be re-optimised. Until an edge proves itself going forward it trades at micro size; if it fails, it is not taken at all. Protection against overfitting across thousands of tests.
Risk sizing and compounding
Each trade's size equals a fixed percentage of equity at risk divided by the ATR stop. Every trade risks the same amount, size grows with the account through compounding, and instruments size themselves via live conversion into the account currency.
Crypto edges traded directly
The same edges are validated by walk-forward on crypto (Binance data, honest taker costs) and traded directly on proven statistics rather than by neural guesswork. A single account compounds, starting in paper mode until confirmed.
Fidelity lab: test versus account
Real closed trades are replayed through the model with identical parameters — the stop and target are taken from the trade itself rather than assumed. The gap between what the model promised and what the account delivered exposes a leak even where none was suspected. The first run found a bot whose trades were killed from outside 86% of the time, reaching neither stop nor target.
Portfolio lab
The other labs measure a single strategy in isolation, but money is lost in combination. Here the whole set trades simultaneously on a shared account, testing the values that used to be guesswork: how many positions to hold at once, at what correlation to stop adding another, and whether the daily loss limit helps.
Exits chosen by measurement
The profit-lock threshold and holding limit are fitted to each bot's actual setups rather than assigned. No universal rule exists: for one bot locking hurts at every level, for another it helps, for a third it makes no difference — only measurement decides.
Risk scale separate from timeframe
Entry timing and trade weight are different things. The stop can be measured on a higher interval than the signal: on the same data a hard binding produced zero survivors out of 47,682 variants, while separating them produced ten.
Measurement at every calculation step
Calculation errors look entirely plausible: the code runs, numbers come out, the logs stay silent. Every step is checked against a physically possible range and any violation is recorded with examples. A strategy with not one signal across its whole history is flagged as dead.
Error class scanner
Bugs multiply by copy-paste — the same line lives in nine bots at once. Each known class is written up as a check and run across every file, preventing it from arriving again with a new bot.
Computation distributed to external machines
Selection is compute-bound: one instrument on one interval downloads in two seconds and computes in two hundred. Jobs are handed to other machines — only candles go out and numbers come back, there are no broker keys in a job, and a computing machine cannot trade.
Selection six times faster
Each strategy's signal was recomputed from scratch forty-five times in a row — once per target/stop pair and once per exit mode. Removing that redundancy cut a run from 185 seconds to 33 with fully identical results.
The AI learns from edges
The network receives a compressed consensus of every proven edge (grouped by family: trend, mean reversion, breakout, range) and learns to weight them by market regime. It stays in shadow behind the AI gate until it proves an edge going forward.
A library of documented edges
We codify validated edges from the research — TSMOM, Connors' RSI(2)/IBS, the Fisher Transform, TRIX, Aroon, the Relative Vigor Index, the Schaff Trend Cycle, TD Sequential, the Hull moving average, Crabel's NR7 — and put them through our own lab. Whatever fails walk-forward is discarded honestly.
Exit ratchets tuned per edge
Profit often comes from the exit, not the entry. The lab picks the best ratchet for each edge individually: Chandelier, break-even after +1R, Donchian/Turtle, an EMA trail or a fixed target — and the executor manages the trade with that trailing method.
An extended universe
Dozens of CFD, FX and crypto markets in one lab: global indices (S&P, Nasdaq, Russell, DAX, Nikkei, FTSE), metals (gold, silver, platinum, copper), energy, soft commodities (wheat, corn, sugar) and major crypto — more markets means more independent edges.
Quantitative core
Levels treated as probabilistic objects (probability of rejection or break, quality), P&L distributions with their tails (value-at-risk, skew, kurtosis), market memory (Hurst) and signal entropy — Quant core →
Risk doctrine 2.0 (tail-aware)
Trade size adapts to regime and tails: a poor regime or a fat value-at-risk means a smaller size. Plus a daily −6% and weekly −12% stop and a hard correlation limit across every executor.
Real microstructure
A genuine crypto order book (Binance depth plus aggressors from the trade tape): book imbalance, who is hitting the market, cumulative delta, spread. A chart widget plus a protective size trim in the bot. To be clear: crypto only — CFD and FX are over-the-counter and have no book.
Structural probability
A logistic regression on structural features yields a calibrated probability of success for an entry — a probability, not a frequency. Calibration is monotone out of sample and training runs weekly. Interpretable, not a black box.
Edge decay detector
Catches a strategy's quiet death: a fall in expectancy across the shadow record triggers an automatic email recommending withdrawal and recalibration. Protection against a slow bleed.
Signal ensemble: entry and exit
Not “one edge, one signal”, but a composite confidence from every edge on the instrument, weighted by quality (lower-bound profit factor) and fit to the market regime. An agreement gate and a size multiplier on entry; when the whole pool turns against the position, it closes without waiting for the stop. The combination governs the entire trade.
Exit bar scan
The exit checks every bar since entry, not just the last closed one. If a bar is missed (a restart, a timeout), the target or stop on that bar is no longer lost — the false bias of closures towards losses is gone. Across every bot.
Priority to runners
Size is weighted by strategy type: trend runners (payoff above 1) get full size, while pure mean reversion (which clips winners) has its risk cut by 40%. A trend filter on entry also prevents fading a strong trend.
Live bot dashboard
One panel: the OANDA account, every executor (crypto / Kraken / Trading 212 / Alpaca), the GTE family and the legends — win rate, payoff, net and on/off switches, refreshed every 20 seconds.