From raw market noise to a tradable decision stack.
The strategy concept treats research as a layered system: collect data, model higher-timeframe opportunity, size from probability, and execute with lower-timeframe triggers that stay simple enough to debug.
Project Objective
Build a reusable research pipeline rather than one isolated model or backtest. The model predicts higher-timeframe opportunity, the sizing layer converts probability into exposure, and the execution layer handles lower-timeframe entries and exits.
Strategy Workflow
Prediction Targets
Direction
Probability that the asset moves up or down over the forecast horizon.
Rate of Return
Expected forward return over the forecast horizon.
Rank
Cross-sectional ordering of assets by expected opportunity.
Probability and Kelly Sizing
Probability becomes exposure
The model should output probabilities or confidence scores. With a fixed take-profit and fixed stop-loss, the payoff ratio is measurable, so the Kelly fraction can be estimated and then capped for risk control.
If confidence looks strong but realized payoff does not support the size, the problem may be calibration, the TP/SL shape, or the target itself.
Multi-Timeframe Design
Higher Timeframe
Daily, H1, or block-level signals predict trend, regime, or opportunity.
Lower Timeframe
Execution waits for concrete triggers such as MA crossover, breakout, pullback, or mean reversion.
CryptoQuant Expansion
Daily, H1, and block data can be compared to test which horizon adds durable edge.
Why Trend Still Needs Entry
Trend says the trade is worth considering
The predictive model works at the higher timeframe. It estimates whether the environment favors a long, a short, or no position at all through three lenses: direction, rate of return, and rank.
This is the opportunity layer. It tells us where the edge may live, but not the exact moment to spend risk.
Entry says whether now is a good moment
The entry layer exists because a correct trend forecast can still be badly timed. A bullish model can still fire while panic, liquidation, or cascade selling is underway. The trend may be right, but entry still matters because we do not want to catch falling knives bare-handed, and we do not want to stand against a rocket.
Entry logic is the timing layer. It reduces the chance of entering too early, too stretched, or into adverse momentum.
A good entry pattern cannot rescue a model with no predictive edge. And a good model can still produce weak realized trades if execution keeps stepping in too early. That is why trend prediction and entry logic should be tested and improved as separate pieces of the stack.
How Results Improve
Expand market, macro, trend, on-chain, multi-timeframe, and regime coverage so the model sees more ways the market can behave.
Build better rolling changes, spreads, interactions, normalization, cross-timeframe features, and regime-aware transformations.
Improve lower-timeframe timing, confirmation, TP/SL shape, and payoff structure so the edge turns into cleaner realized trades.
Upgrade modeling, calibration, ranking objectives, ensembles, regime-aware methods, and validation discipline.
More data, better features, and new modeling techniques mainly strengthen the higher-timeframe prediction layer.
Entry design and risk/reward work mainly improve execution quality by filtering poor timing and reshaping payoff.
OOS quality, PnL quality, drawdown, and stability decide which lever deserves the next research cycle.
If the edge is weak, loop back to the weakest layer instead of adding complexity everywhere at once.
Usually points back to more data, better features, or stronger modeling assumptions.
Usually points back to entry timing, confirmation, and risk/reward design.
Often means the TP/SL assumptions, Kelly cap, or regime filter need work.
Compare daily, H1, and block-level signals before trusting the strategy.
Trend vs Entry
Trend Layer
The trend layer is where the predictive model earns its keep. It tries to estimate whether the market environment supports a trade through direction, forward return, and cross-sectional rank.
This layer is where more data, feature engineering, and new data science techniques usually matter most.
Entry Layer
The entry layer is where execution variants live: MA crossover, breakout, pullback, continuation, confirmation filters, and mean-reversion timing. The goal is not to invent a second prediction model. The goal is to enter in a way that respects momentum, structure, and payoff.
This layer is where timing, TP/SL design, and risk/reward tuning can materially improve realized returns without changing the core trend view.
OOS Validation
Out-of-sample validation is the first overfitting check. Current monthly BTC experiments use chronological train, test, and OOS splits, with test beginning at 2022-12-31, OOS beginning at 2025-01-01, and a 30-day embargo between split boundaries.
The selected model is chosen using train/test, then refit on train plus test, and finally evaluated once on the untouched OOS period. If test works but OOS collapses, the strategy should return to feature engineering, target construction, timeframe selection, or probability calibration.
Do Not Celebrate Too Early
Good returns can be a warning sign
A beautiful backtest is not automatically a good strategy. It may be real edge, but it may also be leakage, survivorship bias, accidental future information, target contamination, bad timestamp alignment, or a methodology mistake that makes the past look tradable.
The right reaction to strong results is curiosity first, celebration later.
When we cannot find the flaw
If the methodology has been checked and we still cannot identify what is wrong, the next step is not to size aggressively. The next step is a controlled live test or paper/live shadow run, so the market can reveal whether the edge survives outside the research environment.
Live testing turns an impressive backtest into a measurable hypothesis.
Failure Loops
- Bad data: return to collection, add missing regimes, and improve alignment before changing the model.
- Weak model metrics: revisit features, labels, target horizon, and modeling technique.
- Weak OOS or unstable edge: usually means more data, better features, or stronger validation discipline are needed.
- Good trend but poor realized trades: redesign the entry method instead of assuming the predictive model is useless.
- High drawdown or poor payoff: rework TP/SL assumptions, Kelly caps, and risk/reward structure.
- Overtrading in bad momentum: add stronger confirmation so the entry layer does not keep stepping into adverse flow.