Adaptive Drill-Down: Backtest with Variable Granularity from Minutes to Milliseconds

Minute candles are the standard granularity for backtests. But within a single minute candle, price can move differently: sometimes by 0.01%, other times by 2%. When both stop-loss and take-profit fall within the [low, high] range of a single minute candle, the backtest doesn't know which triggered first. This is the fill ambiguity problem.

The naive solution is to switch to second-level data for the entire backtest. But over two years, that's ~63 million second bars instead of ~1 million minute bars. Storage increases 60x, speed drops proportionally.

Adaptive drill-down solves this problem: use fine granularity only where it's actually needed.

The Problem: Fill Ambiguity on Large Candles

Consider a specific situation. The strategy opened a long at 3000 USDT. Stop-loss: 2970 (-1%). Take-profit: 3060 (+2%).

The minute candle at 14:37:

Open: 3010
High: 3065
Low: 2965
Close: 3050

Both SL (2970) and TP (3060) fall within the range [2965, 3065]. Which triggered first?

Possible outcomes:

Price went down first -> SL triggered -> loss of -1%
Price went up first -> TP triggered -> profit of +2%

The difference in a single trade: 3 percentage points. With 10x leverage — 30%. For a backtest with hundreds of trades, incorrect fill ambiguity resolution systematically distorts results.

How Frameworks Handle This by Default

Most backtest engines use one of two heuristics:

Optimistic: TP triggers first -> inflated results
Pessimistic: SL triggers first -> deflated results

Both approaches are guesswork. Real data is available at second or even millisecond level, and there's no reason to guess when you can look.

Drill-Down: Three-Level Strategy

Adaptive three-level drill-down resolution pyramid

The drill-down idea: start at the minute level and "drill down" to a lower level only when there's ambiguity.

Level 1: 1m (minute candles)
  -> If SL or TP is unambiguously outside the [low, high] range — resolve on the spot
  -> If both are within the range — drill down

Level 2: 1s (second candles)
  -> Load 60 second bars for this minute
  -> Walk through second by second: which triggered first?
  -> If a second bar is also ambiguous — drill down

Level 3: 100ms (millisecond candles)
  -> Load up to 10 bars of 100ms for this second
  -> Resolve the fill at order book level

When Drill-Down Is Not Needed

In 95% of cases, drill-down is not required. Typical scenarios:

Unambiguous SL: candle high doesn't reach TP, low breaks through SL -> SL triggered, no drill-down needed.

Unambiguous TP: low doesn't reach SL, high breaks through TP -> TP triggered, no drill-down needed.

Neither triggered: both levels are outside the range -> position remains open.

Gap detection: the open of the next candle jumps through SL or TP -> execution at open price, no drill-down.

Drill-down is needed only for ~5% of bars — when both levels fall within the range of a single candle.

class AdaptiveFillSimulator:
    """
    Three-level drill-down for determining fill order.
    """
    def __init__(self, data_loader):
        self.loader = data_loader
        self.cache_1s = {}  # Cache of second data by month

    def check_fill(self, timestamp, candle_1m, sl_price, tp_price, side):
        """
        Checks whether SL or TP triggered on the given minute candle.

        Returns: ('sl', fill_price) | ('tp', fill_price) | None
        """
        low, high = candle_1m['low'], candle_1m['high']

        open_price = candle_1m['open']
        if side == 'long':
            if open_price <= sl_price:
                return ('sl', open_price)
            if open_price >= tp_price:
                return ('tp', open_price)
        else:
            if open_price >= sl_price:
                return ('sl', open_price)
            if open_price <= tp_price:
                return ('tp', open_price)

        sl_hit = self._level_hit(sl_price, low, high, side, 'sl')
        tp_hit = self._level_hit(tp_price, low, high, side, 'tp')

        if sl_hit and not tp_hit:
            return ('sl', sl_price)
        if tp_hit and not sl_hit:
            return ('tp', tp_price)
        if not sl_hit and not tp_hit:
            return None

        return self._drill_down_1s(timestamp, sl_price, tp_price, side)

    def _drill_down_1s(self, minute_ts, sl_price, tp_price, side):
        """Level 2: second-by-second pass."""
        bars_1s = self.loader.load_1s_for_minute(minute_ts)

        if bars_1s is None or len(bars_1s) == 0:
            return self._pessimistic_fill(side, sl_price, tp_price)

        for bar in bars_1s:
            sl_hit = self._level_hit(sl_price, bar['low'], bar['high'], side, 'sl')
            tp_hit = self._level_hit(tp_price, bar['low'], bar['high'], side, 'tp')

            if sl_hit and not tp_hit:
                return ('sl', sl_price)
            if tp_hit and not sl_hit:
                return ('tp', tp_price)
            if sl_hit and tp_hit:
                result = self._drill_down_100ms(bar['timestamp'], sl_price, tp_price, side)
                if result:
                    return result

        return self._pessimistic_fill(side, sl_price, tp_price)

    def _pessimistic_fill(self, side, sl_price, tp_price):
        """Pessimistic assumption: SL for longs, TP for shorts."""
        if side == 'long':
            return ('sl', sl_price)
        else:
            return ('sl', sl_price)

Performance

Mode	Time per fill check	When used
1m (no drill-down)	~0ms	~95% of cases
1s drill-down	~5ms (first access to month)	~5% of cases
100ms drill-down	~1ms	<0.5% of cases

Over a 2-year backtest with ~400 trades, drill-down is invoked for approximately 20 candles. Total overhead — less than 1 second for the entire backtest.

Adaptive Data Storage

Drill-down requires second and millisecond data. But storing everything at maximum granularity is impractical:

Granularity	Bars over 2 years	Parquet size
1m	~1.05M	~15 MB
1s	~63M	~550 MB/month
100ms	~630M	~5 GB/month

A complete 1s archive over 2 years is about 13 GB. 100ms — over 100 GB. Storing everything is possible but wasteful, considering that drill-down uses less than 1% of this data.

Hot-Second Detection

Hot-second detection and adaptive storage savings

The key observation: seconds in which price moves significantly represent a small fraction. If price changed by less than 0.1% within a second — there's no point storing the 100ms breakdown for that second.

Hot-second detection: when downloading and processing data, we analyze each second and generate 100ms candles only for "hot" seconds — those where price movement exceeded the threshold.

def process_trades_adaptive(
    trades: pd.DataFrame,
    min_price_change_pct: float = 1.0,
) -> tuple[pd.DataFrame, pd.DataFrame]:
    """
    Processes raw trades into an adaptive structure:
    - 1s candles for all seconds
    - 100ms candles only for "hot" seconds

    Args:
        trades: DataFrame with columns [timestamp, price, quantity]
        min_price_change_pct: threshold for drill-down to 100ms

    Returns:
        (df_1s, df_100ms_hot) — second candles and 100ms for hot seconds
    """
    trades['second'] = trades['timestamp'].dt.floor('1s')
    df_1s = trades.groupby('second').agg(
        open=('price', 'first'),
        high=('price', 'max'),
        low=('price', 'min'),
        close=('price', 'last'),
        volume=('quantity', 'sum'),
    )

    df_1s['price_change_pct'] = (df_1s['high'] - df_1s['low']) / df_1s['open'] * 100
    hot_seconds = df_1s[df_1s['price_change_pct'] >= min_price_change_pct].index

    hot_trades = trades[trades['second'].isin(hot_seconds)]
    hot_trades['bucket_100ms'] = hot_trades['timestamp'].dt.floor('100ms')

    df_100ms = hot_trades.groupby('bucket_100ms').agg(
        open=('price', 'first'),
        high=('price', 'max'),
        low=('price', 'min'),
        close=('price', 'last'),
        volume=('quantity', 'sum'),
    )

    return df_1s, df_100ms

Storage Savings

For example — ETHUSDT over a typical month:

Approach	Size	Granularity
1m only	~1 MB	1 minute
All 1s	~550 MB	1 second
All 100ms	~5 GB	100 ms
Adaptive	~600 MB	1s + 100ms only for hot seconds

With a threshold of min_price_change_pct = 1.0%, hot seconds account for less than 1% of all seconds. 100ms data for them adds ~50 MB to the 550 MB of second data — a negligible overhead.

If second data is also stored adaptively (only when movement within a minute exceeds 0.1%), the volume can be reduced by another 3-5x.

Parquet Storage Structure

data/{SYMBOL}/
├── klines_1m/
│   ├── 2024-01.parquet       # ~1 MB
│   ├── 2024-02.parquet
│   └── ...
├── klines_1s/
│   ├── 2024-01.parquet       # ~550 MB
│   └── ...
├── klines_100ms_hot/
│   ├── 2024-01.parquet       # ~50 MB (hot seconds only)
│   └── ...
└── states_1m.parquet          # Precomputed rolling state cache (~112 MB)

Each file covers one month of data. Second and millisecond data is loaded lazily — only when drill-down requests it.

Parquet Optimization for Financial Data

Financial data has specific characteristics: timestamps grow monotonically, prices change smoothly, volumes vary significantly. Optimal settings:

import pyarrow as pa
import pyarrow.parquet as pq

schema = pa.schema([
    pa.field("timestamp", pa.int32()),    # Seconds from epoch — int32 is sufficient
    pa.field("open",      pa.float32()),
    pa.field("high",      pa.float32()),
    pa.field("low",       pa.float32()),
    pa.field("close",     pa.float32()),
    pa.field("volume",    pa.float32()),
])

column_encodings = {
    "timestamp": "DELTA_BINARY_PACKED",   # Monotonic int -> delta compression
    "open":      "BYTE_STREAM_SPLIT",     # Float -> byte-stream split
    "high":      "BYTE_STREAM_SPLIT",
    "low":       "BYTE_STREAM_SPLIT",
    "close":     "BYTE_STREAM_SPLIT",
    "volume":    "BYTE_STREAM_SPLIT",
}

def save_optimized_parquet(df, path):
    table = pa.Table.from_pandas(df, schema=schema)
    pq.write_table(
        table, path,
        compression="zstd",
        compression_level=9,
        use_dictionary=False,
        write_statistics=False,
        column_encoding=column_encodings,
    )

Why these settings:

DELTA_BINARY_PACKED for timestamps: consecutive timestamps differ by a fixed value (60 for 1m, 1 for 1s). Delta encoding compresses them to nearly zero.
BYTE_STREAM_SPLIT for float: splits float32 bytes into streams (all first bytes together, all second bytes together, etc.). For smoothly changing prices, this achieves 2-3x better compression than standard encoding.
ZSTD level 9: good compression with acceptable decompression speed.
float32 instead of float64: sufficient for prices and volumes, saves 50% memory.

Lazy Loading with Caching

Drill-down requests second data for a specific minute. Loading a parquet file for each request is slow. The solution — lazy loading with an LRU cache by month.

from functools import lru_cache
import pyarrow.parquet as pq
import pandas as pd

class AdaptiveDataLoader:
    """
    Lazy loader with cache: loads second data by month,
    keeps the last N months in memory.
    """
    def __init__(self, symbol: str, data_dir: str = "data", cache_months: int = 2):
        self.symbol = symbol
        self.data_dir = data_dir
        self.cache_months = cache_months
        self._cache_1s: dict[str, pd.DataFrame] = {}

    def load_1s_for_minute(self, minute_ts: pd.Timestamp) -> pd.DataFrame | None:
        """Load 1s data for a specific minute."""
        month_key = minute_ts.strftime("%Y-%m")

        if month_key not in self._cache_1s:
            self._load_month_1s(month_key)

        if month_key not in self._cache_1s:
            return None

        df = self._cache_1s[month_key]
        minute_start = minute_ts.floor('1min')
        minute_end = minute_start + pd.Timedelta(minutes=1)

        return df[(df.index >= minute_start) & (df.index < minute_end)]

    def load_100ms_for_second(self, second_ts: pd.Timestamp) -> pd.DataFrame | None:
        """Load 100ms data for a hot second."""
        month_key = second_ts.strftime("%Y-%m")
        path = f"{self.data_dir}/{self.symbol}/klines_100ms_hot/{month_key}.parquet"

        try:
            df = pd.read_parquet(path)
            second_start = second_ts.floor('1s')
            second_end = second_start + pd.Timedelta(seconds=1)
            return df[(df.index >= second_start) & (df.index < second_end)]
        except FileNotFoundError:
            return None

    def _load_month_1s(self, month_key: str):
        """Load a month of 1s data, evict old data from cache."""
        path = f"{self.data_dir}/{self.symbol}/klines_1s/{month_key}.parquet"
        try:
            df = pd.read_parquet(path)
            df.index = pd.to_datetime(df['timestamp'], unit='s')

            if len(self._cache_1s) >= self.cache_months:
                oldest = min(self._cache_1s.keys())
                del self._cache_1s[oldest]

            self._cache_1s[month_key] = df
        except FileNotFoundError:
            pass

Applying Drill-Down to Backtesting

Integration into the backtest loop:

def backtest_with_adaptive_fill(
    states: pd.DataFrame,
    strategy_params: dict,
    data_loader: AdaptiveDataLoader,
) -> list:
    """
    Backtest with adaptive drill-down for fill simulation.
    """
    fill_sim = AdaptiveFillSimulator(data_loader)
    trades = []
    position = None

    for i in range(len(states)):
        row = states.iloc[i]
        ts = states.index[i]

        candle_1m = {
            'open': row['open'], 'high': row['high'],
            'low': row['low'], 'close': row['close'],
            'timestamp': ts,
        }

        if position is not None:
            fill = fill_sim.check_fill(
                ts, candle_1m,
                position['sl'], position['tp'],
                position['side'],
            )

            if fill is not None:
                fill_type, fill_price = fill
                trades.append({
                    'entry_time': position['entry_time'],
                    'exit_time': ts,
                    'side': position['side'],
                    'entry_price': position['entry_price'],
                    'exit_price': fill_price,
                    'exit_type': fill_type,
                    'drill_down': fill_sim.last_drill_depth,  # 0, 1, or 2
                })
                position = None
                continue

        signal = check_entry_signal(row, strategy_params)
        if signal and position is None:
            position = {
                'side': signal['side'],
                'entry_price': row['close'],
                'entry_time': ts,
                'sl': signal['sl'],
                'tp': signal['tp'],
            }

    return trades

Relationship with Rolling State Cache

Drill-down complements the aggregated parquet cache — they solve different problems:

	Rolling state cache	Adaptive drill-down
Purpose	Correct HTF indicator values	Precise SL/TP execution order
Operates on	Every 1m candle	Only during fill ambiguity (~5%)
Data	Precomputed, stored permanently	Lazy loaded, cache of recent months
Affects	Entry/exit signals	Execution price and time

Both approaches eliminate errors invisible at the daily candle level but critical for realistic backtesting.

Summary: Fill Simulation Approach Comparison

Approach	Accuracy	Speed	Storage
OHLC heuristic (optimist/pessimist)	Low	Instant	1m only
Full 1s backtest	High	Slow (x60)	~550 MB/month
Full 100ms backtest	Maximum	Very slow (x600)	~5 GB/month
Adaptive drill-down	High	~Instant	1m + 1s + 100ms hot

Drill-down provides the accuracy of a full 1s backtest at the speed of a 1m backtest. The key observation: high granularity is not needed everywhere — only at decision points.

Conclusion

Adaptive drill-down is the application of a simple principle: spend computational resources and storage proportionally to data importance.

Three granularity levels:

1m — base pass for 95% of bars
1s — drill-down during fill ambiguity
100ms — drill-down for hot seconds with extreme movement

Three storage levels:

All 1m — complete archive, ~15 MB for 2 years
All 1s — complete or adaptive archive, ~550 MB/month
Hot 100ms only — <1% of seconds, ~50 MB/month

The result: a backtest with tick simulator accuracy at minute-level speed. And storage that grows linearly, not exponentially, with increasing granularity.

For more on precomputed cache for multi-timeframe strategies, see the article Aggregated Parquet Cache. On the impact of funding rates on results with high leverage — Funding rates kill your leverage.

Useful Links

Citation

@article{soloviov2026adaptivedrilldown,
  author = {Soloviov, Eugen},
  title = {Adaptive Drill-Down: Backtest with Variable Granularity from Minutes to Milliseconds},
  year = {2026},
  url = {https://marketmaker.cc/ru/blog/post/adaptive-resolution-drill-down-backtest},
  description = {How adaptive data granularity speeds up backtests and saves storage: drill-down from 1m to 1s and 100ms only where price moved significantly.}
}

Adaptive drill-down solves this problem: use fine granularity only where it's actually needed.

The Problem: Fill Ambiguity on Large Candles

Consider a specific situation. The strategy opened a long at 3000 USDT. Stop-loss: 2970 (-1%). Take-profit: 3060 (+2%).

The minute candle at 14:37:

Open: 3010
High: 3065
Low: 2965
Close: 3050

Both SL (2970) and TP (3060) fall within the range [2965, 3065]. Which triggered first?

Possible outcomes:

Price went down first -> SL triggered -> loss of -1%
Price went up first -> TP triggered -> profit of +2%

The difference in a single trade: 3 percentage points. With 10x leverage — 30%. For a backtest with hundreds of trades, incorrect fill ambiguity resolution systematically distorts results.

How Frameworks Handle This by Default

Most backtest engines use one of two heuristics:

Optimistic: TP triggers first -> inflated results
Pessimistic: SL triggers first -> deflated results

Both approaches are guesswork. Real data is available at second or even millisecond level, and there's no reason to guess when you can look.

Drill-Down: Three-Level Strategy

Adaptive three-level drill-down resolution pyramid

The drill-down idea: start at the minute level and "drill down" to a lower level only when there's ambiguity.

Level 1: 1m (minute candles)
  -> If SL or TP is unambiguously outside the [low, high] range — resolve on the spot
  -> If both are within the range — drill down

Level 2: 1s (second candles)
  -> Load 60 second bars for this minute
  -> Walk through second by second: which triggered first?
  -> If a second bar is also ambiguous — drill down

Level 3: 100ms (millisecond candles)
  -> Load up to 10 bars of 100ms for this second
  -> Resolve the fill at order book level

When Drill-Down Is Not Needed

In 95% of cases, drill-down is not required. Typical scenarios:

Unambiguous SL: candle high doesn't reach TP, low breaks through SL -> SL triggered, no drill-down needed.

Unambiguous TP: low doesn't reach SL, high breaks through TP -> TP triggered, no drill-down needed.

Neither triggered: both levels are outside the range -> position remains open.

Gap detection: the open of the next candle jumps through SL or TP -> execution at open price, no drill-down.

Drill-down is needed only for ~5% of bars — when both levels fall within the range of a single candle.

class AdaptiveFillSimulator:
    """
    Three-level drill-down for determining fill order.
    """
    def __init__(self, data_loader):
        self.loader = data_loader
        self.cache_1s = {}  # Cache of second data by month

    def check_fill(self, timestamp, candle_1m, sl_price, tp_price, side):
        """
        Checks whether SL or TP triggered on the given minute candle.

        Returns: ('sl', fill_price) | ('tp', fill_price) | None
        """
        low, high = candle_1m['low'], candle_1m['high']

        open_price = candle_1m['open']
        if side == 'long':
            if open_price <= sl_price:
                return ('sl', open_price)
            if open_price >= tp_price:
                return ('tp', open_price)
        else:
            if open_price >= sl_price:
                return ('sl', open_price)
            if open_price <= tp_price:
                return ('tp', open_price)

        sl_hit = self._level_hit(sl_price, low, high, side, 'sl')
        tp_hit = self._level_hit(tp_price, low, high, side, 'tp')

        if sl_hit and not tp_hit:
            return ('sl', sl_price)
        if tp_hit and not sl_hit:
            return ('tp', tp_price)
        if not sl_hit and not tp_hit:
            return None

        return self._drill_down_1s(timestamp, sl_price, tp_price, side)

    def _drill_down_1s(self, minute_ts, sl_price, tp_price, side):
        """Level 2: second-by-second pass."""
        bars_1s = self.loader.load_1s_for_minute(minute_ts)

        if bars_1s is None or len(bars_1s) == 0:
            return self._pessimistic_fill(side, sl_price, tp_price)

        for bar in bars_1s:
            sl_hit = self._level_hit(sl_price, bar['low'], bar['high'], side, 'sl')
            tp_hit = self._level_hit(tp_price, bar['low'], bar['high'], side, 'tp')

            if sl_hit and not tp_hit:
                return ('sl', sl_price)
            if tp_hit and not sl_hit:
                return ('tp', tp_price)
            if sl_hit and tp_hit:
                result = self._drill_down_100ms(bar['timestamp'], sl_price, tp_price, side)
                if result:
                    return result

        return self._pessimistic_fill(side, sl_price, tp_price)

    def _pessimistic_fill(self, side, sl_price, tp_price):
        """Pessimistic assumption: SL for longs, TP for shorts."""
        if side == 'long':
            return ('sl', sl_price)
        else:
            return ('sl', sl_price)

Performance

Mode	Time per fill check	When used
1m (no drill-down)	~0ms	~95% of cases
1s drill-down	~5ms (first access to month)	~5% of cases
100ms drill-down	~1ms	<0.5% of cases

Over a 2-year backtest with ~400 trades, drill-down is invoked for approximately 20 candles. Total overhead — less than 1 second for the entire backtest.

Adaptive Data Storage

Drill-down requires second and millisecond data. But storing everything at maximum granularity is impractical:

Granularity	Bars over 2 years	Parquet size
1m	~1.05M	~15 MB
1s	~63M	~550 MB/month
100ms	~630M	~5 GB/month

A complete 1s archive over 2 years is about 13 GB. 100ms — over 100 GB. Storing everything is possible but wasteful, considering that drill-down uses less than 1% of this data.

Hot-Second Detection

Hot-second detection and adaptive storage savings

Hot-second detection: when downloading and processing data, we analyze each second and generate 100ms candles only for "hot" seconds — those where price movement exceeded the threshold.

def process_trades_adaptive(
    trades: pd.DataFrame,
    min_price_change_pct: float = 1.0,
) -> tuple[pd.DataFrame, pd.DataFrame]:
    """
    Processes raw trades into an adaptive structure:
    - 1s candles for all seconds
    - 100ms candles only for "hot" seconds

    Args:
        trades: DataFrame with columns [timestamp, price, quantity]
        min_price_change_pct: threshold for drill-down to 100ms

    Returns:
        (df_1s, df_100ms_hot) — second candles and 100ms for hot seconds
    """
    trades['second'] = trades['timestamp'].dt.floor('1s')
    df_1s = trades.groupby('second').agg(
        open=('price', 'first'),
        high=('price', 'max'),
        low=('price', 'min'),
        close=('price', 'last'),
        volume=('quantity', 'sum'),
    )

    df_1s['price_change_pct'] = (df_1s['high'] - df_1s['low']) / df_1s['open'] * 100
    hot_seconds = df_1s[df_1s['price_change_pct'] >= min_price_change_pct].index

    hot_trades = trades[trades['second'].isin(hot_seconds)]
    hot_trades['bucket_100ms'] = hot_trades['timestamp'].dt.floor('100ms')

    df_100ms = hot_trades.groupby('bucket_100ms').agg(
        open=('price', 'first'),
        high=('price', 'max'),
        low=('price', 'min'),
        close=('price', 'last'),
        volume=('quantity', 'sum'),
    )

    return df_1s, df_100ms

Storage Savings

For example — ETHUSDT over a typical month:

Approach	Size	Granularity
1m only	~1 MB	1 minute
All 1s	~550 MB	1 second
All 100ms	~5 GB	100 ms
Adaptive	~600 MB	1s + 100ms only for hot seconds

With a threshold of min_price_change_pct = 1.0%, hot seconds account for less than 1% of all seconds. 100ms data for them adds ~50 MB to the 550 MB of second data — a negligible overhead.

If second data is also stored adaptively (only when movement within a minute exceeds 0.1%), the volume can be reduced by another 3-5x.

Parquet Storage Structure

data/{SYMBOL}/
├── klines_1m/
│   ├── 2024-01.parquet       # ~1 MB
│   ├── 2024-02.parquet
│   └── ...
├── klines_1s/
│   ├── 2024-01.parquet       # ~550 MB
│   └── ...
├── klines_100ms_hot/
│   ├── 2024-01.parquet       # ~50 MB (hot seconds only)
│   └── ...
└── states_1m.parquet          # Precomputed rolling state cache (~112 MB)

Each file covers one month of data. Second and millisecond data is loaded lazily — only when drill-down requests it.

Parquet Optimization for Financial Data

Financial data has specific characteristics: timestamps grow monotonically, prices change smoothly, volumes vary significantly. Optimal settings:

import pyarrow as pa
import pyarrow.parquet as pq

schema = pa.schema([
    pa.field("timestamp", pa.int32()),    # Seconds from epoch — int32 is sufficient
    pa.field("open",      pa.float32()),
    pa.field("high",      pa.float32()),
    pa.field("low",       pa.float32()),
    pa.field("close",     pa.float32()),
    pa.field("volume",    pa.float32()),
])

column_encodings = {
    "timestamp": "DELTA_BINARY_PACKED",   # Monotonic int -> delta compression
    "open":      "BYTE_STREAM_SPLIT",     # Float -> byte-stream split
    "high":      "BYTE_STREAM_SPLIT",
    "low":       "BYTE_STREAM_SPLIT",
    "close":     "BYTE_STREAM_SPLIT",
    "volume":    "BYTE_STREAM_SPLIT",
}

def save_optimized_parquet(df, path):
    table = pa.Table.from_pandas(df, schema=schema)
    pq.write_table(
        table, path,
        compression="zstd",
        compression_level=9,
        use_dictionary=False,
        write_statistics=False,
        column_encoding=column_encodings,
    )

Why these settings:

DELTA_BINARY_PACKED for timestamps: consecutive timestamps differ by a fixed value (60 for 1m, 1 for 1s). Delta encoding compresses them to nearly zero.
BYTE_STREAM_SPLIT for float: splits float32 bytes into streams (all first bytes together, all second bytes together, etc.). For smoothly changing prices, this achieves 2-3x better compression than standard encoding.
ZSTD level 9: good compression with acceptable decompression speed.
float32 instead of float64: sufficient for prices and volumes, saves 50% memory.

Lazy Loading with Caching

Drill-down requests second data for a specific minute. Loading a parquet file for each request is slow. The solution — lazy loading with an LRU cache by month.

from functools import lru_cache
import pyarrow.parquet as pq
import pandas as pd

class AdaptiveDataLoader:
    """
    Lazy loader with cache: loads second data by month,
    keeps the last N months in memory.
    """
    def __init__(self, symbol: str, data_dir: str = "data", cache_months: int = 2):
        self.symbol = symbol
        self.data_dir = data_dir
        self.cache_months = cache_months
        self._cache_1s: dict[str, pd.DataFrame] = {}

    def load_1s_for_minute(self, minute_ts: pd.Timestamp) -> pd.DataFrame | None:
        """Load 1s data for a specific minute."""
        month_key = minute_ts.strftime("%Y-%m")

        if month_key not in self._cache_1s:
            self._load_month_1s(month_key)

        if month_key not in self._cache_1s:
            return None

        df = self._cache_1s[month_key]
        minute_start = minute_ts.floor('1min')
        minute_end = minute_start + pd.Timedelta(minutes=1)

        return df[(df.index >= minute_start) & (df.index < minute_end)]

    def load_100ms_for_second(self, second_ts: pd.Timestamp) -> pd.DataFrame | None:
        """Load 100ms data for a hot second."""
        month_key = second_ts.strftime("%Y-%m")
        path = f"{self.data_dir}/{self.symbol}/klines_100ms_hot/{month_key}.parquet"

        try:
            df = pd.read_parquet(path)
            second_start = second_ts.floor('1s')
            second_end = second_start + pd.Timedelta(seconds=1)
            return df[(df.index >= second_start) & (df.index < second_end)]
        except FileNotFoundError:
            return None

    def _load_month_1s(self, month_key: str):
        """Load a month of 1s data, evict old data from cache."""
        path = f"{self.data_dir}/{self.symbol}/klines_1s/{month_key}.parquet"
        try:
            df = pd.read_parquet(path)
            df.index = pd.to_datetime(df['timestamp'], unit='s')

            if len(self._cache_1s) >= self.cache_months:
                oldest = min(self._cache_1s.keys())
                del self._cache_1s[oldest]

            self._cache_1s[month_key] = df
        except FileNotFoundError:
            pass

Applying Drill-Down to Backtesting

Integration into the backtest loop:

def backtest_with_adaptive_fill(
    states: pd.DataFrame,
    strategy_params: dict,
    data_loader: AdaptiveDataLoader,
) -> list:
    """
    Backtest with adaptive drill-down for fill simulation.
    """
    fill_sim = AdaptiveFillSimulator(data_loader)
    trades = []
    position = None

    for i in range(len(states)):
        row = states.iloc[i]
        ts = states.index[i]

        candle_1m = {
            'open': row['open'], 'high': row['high'],
            'low': row['low'], 'close': row['close'],
            'timestamp': ts,
        }

        if position is not None:
            fill = fill_sim.check_fill(
                ts, candle_1m,
                position['sl'], position['tp'],
                position['side'],
            )

            if fill is not None:
                fill_type, fill_price = fill
                trades.append({
                    'entry_time': position['entry_time'],
                    'exit_time': ts,
                    'side': position['side'],
                    'entry_price': position['entry_price'],
                    'exit_price': fill_price,
                    'exit_type': fill_type,
                    'drill_down': fill_sim.last_drill_depth,  # 0, 1, or 2
                })
                position = None
                continue

        signal = check_entry_signal(row, strategy_params)
        if signal and position is None:
            position = {
                'side': signal['side'],
                'entry_price': row['close'],
                'entry_time': ts,
                'sl': signal['sl'],
                'tp': signal['tp'],
            }

    return trades

Relationship with Rolling State Cache

Drill-down complements the aggregated parquet cache — they solve different problems:

	Rolling state cache	Adaptive drill-down
Purpose	Correct HTF indicator values	Precise SL/TP execution order
Operates on	Every 1m candle	Only during fill ambiguity (~5%)
Data	Precomputed, stored permanently	Lazy loaded, cache of recent months
Affects	Entry/exit signals	Execution price and time

Both approaches eliminate errors invisible at the daily candle level but critical for realistic backtesting.

Summary: Fill Simulation Approach Comparison

Approach	Accuracy	Speed	Storage
OHLC heuristic (optimist/pessimist)	Low	Instant	1m only
Full 1s backtest	High	Slow (x60)	~550 MB/month
Full 100ms backtest	Maximum	Very slow (x600)	~5 GB/month
Adaptive drill-down	High	~Instant	1m + 1s + 100ms hot

Drill-down provides the accuracy of a full 1s backtest at the speed of a 1m backtest. The key observation: high granularity is not needed everywhere — only at decision points.

Conclusion

Adaptive drill-down is the application of a simple principle: spend computational resources and storage proportionally to data importance.

Three granularity levels:

1m — base pass for 95% of bars
1s — drill-down during fill ambiguity
100ms — drill-down for hot seconds with extreme movement

Three storage levels:

All 1m — complete archive, ~15 MB for 2 years
All 1s — complete or adaptive archive, ~550 MB/month
Hot 100ms only — <1% of seconds, ~50 MB/month

The result: a backtest with tick simulator accuracy at minute-level speed. And storage that grows linearly, not exponentially, with increasing granularity.

Useful Links

Citation

@article{soloviov2026adaptivedrilldown,
  author = {Soloviov, Eugen},
  title = {Adaptive Drill-Down: Backtest with Variable Granularity from Minutes to Milliseconds},
  year = {2026},
  url = {https://marketmaker.cc/ru/blog/post/adaptive-resolution-drill-down-backtest},
  description = {How adaptive data granularity speeds up backtests and saves storage: drill-down from 1m to 1s and 100ms only where price moved significantly.}
}

Adaptive Drill-Down: Backtest with Variable Granularity from Minutes to Milliseconds

The Problem: Fill Ambiguity on Large Candles

How Frameworks Handle This by Default

Drill-Down: Three-Level Strategy

When Drill-Down Is Not Needed

Performance

Adaptive Data Storage

Hot-Second Detection

Storage Savings

Parquet Storage Structure

Parquet Optimization for Financial Data

Lazy Loading with Caching

Applying Drill-Down to Backtesting

Relationship with Rolling State Cache

Summary: Fill Simulation Approach Comparison

Conclusion

Useful Links

Citation

Read More

Aggregated Parquet Cache: How to Speed Up Multi-Timeframe Backtests by Hundreds of Times

Walk-Forward Optimization: The Only Honest Strategy Test

Plateau Analysis: How to Distinguish a Robust Optimum from Overfitting

The Problem: Fill Ambiguity on Large Candles

How Frameworks Handle This by Default

Drill-Down: Three-Level Strategy

When Drill-Down Is Not Needed

Performance

Adaptive Data Storage

Hot-Second Detection

Storage Savings

Parquet Storage Structure

Parquet Optimization for Financial Data

Lazy Loading with Caching

Applying Drill-Down to Backtesting

Relationship with Rolling State Cache

Summary: Fill Simulation Approach Comparison

Conclusion

Useful Links

Citation

The Problem: Fill Ambiguity on Large Candles

How Frameworks Handle This by Default

Drill-Down: Three-Level Strategy

When Drill-Down Is Not Needed

Performance

Adaptive Data Storage

Hot-Second Detection

Storage Savings

Parquet Storage Structure

Parquet Optimization for Financial Data

Lazy Loading with Caching

Applying Drill-Down to Backtesting

Relationship with Rolling State Cache

Summary: Fill Simulation Approach Comparison

Conclusion

Useful Links

Citation

Read More

Aggregated Parquet Cache: How to Speed Up Multi-Timeframe Backtests by Hundreds of Times

Walk-Forward Optimization: The Only Honest Strategy Test

Plateau Analysis: How to Distinguish a Robust Optimum from Overfitting

市場の先を行く

The Problem: Fill Ambiguity on Large Candles

How Frameworks Handle This by Default

Drill-Down: Three-Level Strategy

When Drill-Down Is Not Needed

Performance

Adaptive Data Storage

Hot-Second Detection

Storage Savings

Parquet Storage Structure

Parquet Optimization for Financial Data

Lazy Loading with Caching

Applying Drill-Down to Backtesting

Relationship with Rolling State Cache

Summary: Fill Simulation Approach Comparison

Conclusion

Useful Links

Citation