量化研究引擎，助您构建交易优势。

@vbt.cv_split(
    splitter="from_rolling",
    splitter_kwargs=dict(length=360, split=270, offset=-90, offset_anchor_set=None,
                         backwards="sorted", set_labels=["train", "test"]),
    index_from="data", merge_func="concat", return_grid="all",
)
def sma_sharpe(data, range_, fast, slow):
    start = max(range_.start - slow - 1, 0)
    close = data.close.iloc[start:range_.stop]
    fast_ma = close.vbt.rolling_mean(fast)
    slow_ma = close.vbt.rolling_mean(slow)
    warmup = range_.start - start
    pf = vbt.PF.from_signals(
        data.iloc[range_],
        entries=fast_ma.vbt.crossed_above(slow_ma).vbt.signals.fshift().iloc[warmup:],
        exits=fast_ma.vbt.crossed_below(slow_ma).vbt.signals.fshift().iloc[warmup:],
        fees=0.001,
    )
    return pf.sharpe_ratio

data = vbt.YFData.pull("BTC-USD", start="2020-01-01", end="2026-10-01")
windows = np.arange(5, 105, 5)
grid, best = sma_sharpe(
    data, vbt.Rep("range_"),
    vbt.Param(windows, condition="fast < slow"),
    vbt.Param(windows),
)

~/research $ vbt agents install

  VectorBT PRO v2026.10.5  ·  Rust v2026.10.5  ·  Python 3.12.11

✓ Installed VBT 2026.10.5 skills into /Users/quant/research

.agents/skills, .claude/skills: vbt-analysis, vbt-backtesting, vbt-basics, vbt-callbacks, vbt-cross-validation,
vbt-custom-simulators, vbt-data, vbt-indicators, vbt-numba, vbt-optimization, vbt-plotting, vbt-portfolio-optimization,
vbt-rust-backend, vbt-rust-native, vbt-signals

~/research $ claude

> Is an SMA crossover on BTC worth trading? Validate it walk-forward with train-only parameter selection.

Skill(vbt-cross-validation)
  Train-only selection with @vbt.cv_split

vbt.get_concept("walk-forward analysis")
  10 sources ranked

vbt.get_source("vbt.cv_split")
  Signature and selection rules confirmed

Write(sma_walk_forward.py)
  @vbt.cv_split(
      splitter="from_rolling",
      splitter_kwargs=dict(length=360, split=270, offset=-90, offset_anchor_set=None,
                           backwards="sorted", set_labels=["train", "test"]),
      index_from="data", merge_func="concat", return_grid="all",
  )
  def sma_sharpe(data, range_, fast, slow):
      start = max(range_.start - slow - 1, 0)
      close = data.close.iloc[start:range_.stop]
      fast_ma = close.vbt.rolling_mean(fast)
      slow_ma = close.vbt.rolling_mean(slow)
      warmup = range_.start - start
      pf = vbt.PF.from_signals(
          data.iloc[range_],
          entries=fast_ma.vbt.crossed_above(slow_ma).vbt.signals.fshift().iloc[warmup:],
          exits=fast_ma.vbt.crossed_below(slow_ma).vbt.signals.fshift().iloc[warmup:],
          fees=0.001,
      )
      return pf.sharpe_ratio

  data = vbt.YFData.pull("BTC-USD", start="2020-01-01", end="2026-10-01")
  windows = np.arange(5, 105, 5)
  grid, best = sma_sharpe(
      data, vbt.Rep("range_"),
      vbt.Param(windows, condition="fast < slow"),
      vbt.Param(windows),
  )

vbt.run_code(sma_walk_forward.py)
  9,120 backtests across 24 rolling splits

                        train   test
  Mean Sharpe            1.85  -0.25
  Selection percentile      -   0.51
  Buy-and-hold Sharpe       -   0.91

Not worth trading. Sharpe falls from 1.85 in-sample to -0.25 out of sample, and the selected windows rank at the 51st
percentile of the grid on unseen data, no better than picking at random.