Montealgos · Algorithmic trading firm
We are a tech driven firm focused on algorithmic trading strategies. Thanks to a deep software engineering background we build the trading software ourselves, and keep complete control over the process from end to end: research, backtesting, optimization, validation, selection, execution, reporting, automation and artificial intelligence.
Approach
Most trading shops buy their stack and rent their edge. We took the opposite route. Every layer — the historical data store, the candle engine, the indicator library, the strategy builder, the optimizer, the validation pipeline, the portfolio allocator and the exchange executers — is written in house, in TypeScript, and tested like production software.
That control is the edge. When a strategy behaves oddly in production, we can read the exact line of code that produced the signal, replay the exact candles that produced the trade, and reproduce the result bit for bit. Nothing is a black box we cannot open.
Rules are declared as data. The same strategy object drives the backtest, the forward test and the live account, so there is no translation gap between research and money.
We never ship the best parameter set. We ship the one sitting in the widest stable plateau, verified by stability scans, permutations and out of sample windows.
A hundred small, uncorrelated, low exposure systems across many pairs and timeframes beats one large conviction bet. Exposure is budgeted per symbol and per direction.
Idea generation, optimization, validation, promotion, execution and monitoring run on schedulers. Humans review the dashboards, they do not place the orders.
The pipeline
Thousands of candidate strategies enter the research pipeline. Each stage is a filter with a hard rejection rule, and a strategy only reaches a live account after surviving every one of them.
An agent composes new strategy candidates from a library of 279 indicators, entry and exit presets, filters and risk templates. Ideas come from research notes, published literature and structured prompting; the output is always a validated strategy object, never free text.
Every candidate runs across the full universe — 25 perpetual pairs by 6 timeframes from 2h to 1d — on candles reaching back to 2017, with fees and slippage applied. Anything under 40 trades, a 1.1 profit factor or above 90% time in market is dropped before optimization is allowed to touch it.
Parameters are optimized in ordered steps — indicator lookbacks first, then filters, then exits and risk — with up to 17 variables per strategy. Grid, chunking and Bayesian optimizers each propose candidates, and the top 100 symbol fits are carried forward.
The optimized parameters meet data they have never seen: a strictly out of sample window that starts the day the in sample period ends. Equity and drawdown are resampled weekly and the strategy is rejected on the spot if the window closes below the threshold.
A second untouched window, closer to the present, replays the strategy against the most recent market regime. Forward and live must both be credible: at least one positive, and the pair positive together, or the candidate is discarded.
Every parameter is nudged around its chosen value and the strategy is re-run on each neighbour. A strategy that only works on one exact setting is a curve fit, not an edge — we keep the ones sitting on a broad, flat plateau of results.
Monte Carlo permutation testing shuffles the price series itself and re-runs the whole optimization on synthetic data. If the strategy scores as well on noise as it does on the real market, its result was luck, and it never reaches a portfolio.
Survivors compete for a seat. Selection balances score against correlation, time in market and symbol concentration, so the portfolio keeps a spread of pairs, timeframes, directions and strategy families rather than a hundred variations of the same trade.
Selected strategies are promoted to live accounts. A scheduler wakes on every candle close, recomputes the signals from the same code path the backtest used, sizes the position from its risk budget and routes the order to the exchange. Fills, exposure and drawdown stream back to the liveboard.
The engine
One TypeScript codebase covers the whole chain, from the first candle downloaded to the last order acknowledged. Here is what lives inside it.
279 indicators implemented and unit tested against reference libraries — oscillators, moving averages, bands, volatility, volume, candlestick patterns and custom research signals — available in instant, streaming and multi timeframe flavours.
Candles from 2017 onward, stored locally and served to the backtester in batches. Higher timeframe series are derived and aligned to the base series, so a strategy can read a daily trend filter while trading a 4h signal without look ahead.
Strategies are declarative objects, not scripts: indicators, signal expressions, filters, presets, exits and risk. Three generations run side by side — classic rule builders and a machine learning builder with KNN classifiers and kernel regression filters.
A staged optimizer that treats parameters by role rather than as one flat search space, with grid, chunking, Bayesian and stability oriented modes, and a scoring function that rewards consistency rather than the single highest return.
Permutation tests on synthetic price series, parameter neighbourhood scans, weekly equity and drawdown resampling, all time high ratios, profit factor and Sharpe — the evidence a strategy has to produce before anyone talks about capital.
Native clients for Bybit, Binance, Bitget and Gate on USDT perpetuals, each with a testnet path, one way position mode, symbol filters and reconciliation, so the same strategy can be routed wherever the liquidity and the funding are best.
Two cron driven engines: the research scheduler grinds the optimization and validation queue continuously, while the live scheduler wakes on candle closes to compute signals, place orders and publish state.
Language models are used where they are actually good: proposing strategy structures against a strict schema, reading indicator documentation through tools, and turning research notes into testable candidates. Every suggestion is then judged by the same pipeline as any other.
The numbers
The current selection holds 134 strategies. Below is how that same portfolio behaves on the data it was fitted on, on the out of sample window that followed, and on the most recent window — the honest three way split every systematic result should be shown with.
→
→
→
103 long systems and 31 short systems, each carrying a small fixed risk. Position limits are set per symbol so no single pair can dominate the book.
Backtest, forward and live window figures are simulated results produced by our own research pipeline over the periods stated, net of fees and slippage assumptions. Simulated results have inherent limitations and are not a reliable indicator of future performance. The curves drawn on the three cards are illustrative shapes, not plotted from trade data. Nothing on this page is investment advice or an offer of any financial product.
The portfolio
The cards above are sketches. This one is not: the 158 strategies of portfolio_ccex1_v1 chained through backtest, forward test and live window, trade by trade. The portfolio page opens it up — window by window, month by month, pair by pair.
The platform
The lab. Thousands of models and optimizations with their full history, equity and drawdown curves for every window, stability and permutation scores, correlation views, selection tooling and portfolio simulation — plus a playground for building and inspecting indicators candle by candle.
The control room. Open PnL, current and maximum drawdown, live exposure per symbol and direction, open trades, full trade history, per system and per model breakdowns, and an alert feed that reports anything the schedulers could not complete.
Contact
Interested in the research, the technology, or a conversation about working together? We answer every serious message.
contact@montealgos.com