How SparklingAI Evaluates Multi-Horizon XAUUSD Alpha Signals

One prediction horizon is rarely enough for XAUUSD research.

Gold can move in short bursts, trend across sessions, and reverse sharply around news or liquidity changes. A model that only asks "what happens next?" may miss the difference between a small short-term wiggle and a larger move that develops over several bars.

SparklingAI is therefore moving toward multi-horizon forecasting for alpha research. The goal is not to publish a magic signal. The goal is to understand whether an AI model can produce useful evidence across different future windows.

Why Multi-Horizon Forecasting Matters

In trading research, a horizon is the future window being studied. A short horizon may ask what happens over the next few hours. A longer horizon may ask whether the market has enough directional pressure to matter over a wider window.

This matters because each horizon answers a different question:

• Short horizons can detect immediate movement but may be noisy.
• Medium horizons can reduce some noise but may react more slowly.
• Longer horizons can reveal stronger directional pressure but require more patience.
• A live agent should understand which horizon a signal belongs to before acting on it.

For SparklingAI, the horizon is part of the signal contract. A signal without a horizon is incomplete.

What SparklingAI Measures Publicly

The private alpha recipe is not public. The public diagnostics can still be useful.

For each research fold, SparklingAI can summarize:

• The test period being evaluated
• The forecasting horizons studied
• Rank correlation between predicted and realized movement
• Direction behavior
• How many signals were selected after public-safe filters
• The realized hit rate of selected signals
• Whether the sample is large enough to treat seriously

These metrics are not the same as live profit and loss. They are model diagnostics. They help decide whether a signal deserves deeper backtesting, execution simulation, or rejection.

A Fold-3 Example From The Latest XAUUSD Diagnostics

The latest XAUUSD diagnostic artifact generated on May 25, 2026 includes a fold-3 test window from May 13, 2025 to May 25, 2026.

In that fold, the longer horizons were more interesting than the shorter horizons:

Horizon	Rank IC	Selected signals	Selected hit rate	Selected average signed return
36h	0.202	0	N/A	N/A
72h	0.196	51	90.2%	0.645%
144h	0.240	48	85.4%	1.524%

This does not prove the model is ready for live trading. It does suggest that the public research should pay attention to longer-horizon alpha behavior, not only short-horizon predictions.

The important part is the discipline: report the horizon, the fold, the selected-signal count, and the limits of the sample.

Why Selected Signal Count Matters

A high hit rate with only a few signals can be misleading. A large sample with weak ranking can also be misleading.

SparklingAI treats selected signal count as part of the interpretation. If a horizon shows good ranking behavior but no selected signals, that is not the same as an active trading signal. If a horizon shows selected signals, the next question is whether the sample survives costs, slippage, drawdown, and risk controls.

That is why alpha diagnostics are only one step in the stack.

What This Means For A Future AI Trading Agent

A future SparklingAI agent should not simply receive "buy" or "sell". It should know the forecast horizon and the confidence context behind that signal.

A better agent-facing output would include:

• The symbol being studied
• The horizon being evaluated
• The direction or no-action state
• A confidence or probability measure
• An uncertainty measure
• A reason the signal was accepted, skipped, or blocked

This is how multi-horizon forecasting connects to a future AI trading signal API. The model output becomes structured evidence, not an automatic command.

Public Research Without Giving Away The Alpha

There is a balance between transparency and protecting the future product.

SparklingAI can share public diagnostics such as horizon behavior, fold windows, selected-signal counts, and research interpretation. It does not need to publish the private feature construction, model weights, thresholds, or execution recipe.

That is the line this website is trying to follow: educational, transparent, and research-focused, without turning the alpha layer into a copyable recipe.

For the model direction behind this work, read why SparklingAI is exploring xLSTM for alpha research. For the next layer of the stack, read how SparklingAI validates XAUUSD alpha signals before execution.